Optimization of Finite Dimensional Structures

K11056_FM.indd 1

6/11/10 3:18:11 PM

K11056_FM.indd 2

6/11/10 3:18:11 PM

Optimization of Finite Dimensional Structures

Makoto Ohsaki Hiroshima University Higashi-Hiroshima, Japan

K11056_FM.indd 3

6/11/10 3:18:11 PM

CRC Press Taylor & Francis Group 6000 Broken Sound Parkway NW, Suite 300 Boca Raton, FL 33487-2742 © 2011 by Taylor and Francis Group, LLC CRC Press is an imprint of Taylor & Francis Group, an Informa business No claim to original U.S. Government works Printed in the United States of America on acid-free paper 10 9 8 7 6 5 4 3 2 1 International Standard Book Number-13: 978-1-4398-2004-9 (Ebook-PDF) This book contains information obtained from authentic and highly regarded sources. Reasonable efforts have been made to publish reliable data and information, but the author and publisher cannot assume responsibility for the validity of all materials or the consequences of their use. The authors and publishers have attempted to trace the copyright holders of all material reproduced in this publication and apologize to copyright holders if permission to publish in this form has not been obtained. If any copyright material has not been acknowledged please write and let us know so we may rectify in any future reprint. Except as permitted under U.S. Copyright Law, no part of this book may be reprinted, reproduced, transmitted, or utilized in any form by any electronic, mechanical, or other means, now known or hereafter invented, including photocopying, microfilming, and recording, or in any information storage or retrieval system, without written permission from the publishers. For permission to photocopy or use material electronically from this work, please access www.copyright. com (http://www.copyright.com/) or contact the Copyright Clearance Center, Inc. (CCC), 222 Rosewood Drive, Danvers, MA 01923, 978-750-8400. CCC is a not-for-profit organization that provides licenses and registration for a variety of users. For organizations that have been granted a photocopy license by the CCC, a separate system of payment has been arranged. Trademark Notice: Product or corporate names may be trademarks or registered trademarks, and are used only for identification and explanation without intent to infringe. Visit the Taylor & Francis Web site at http://www.taylorandfrancis.com and the CRC Press Web site at http://www.crcpress.com

Preface

The attempt to ﬁnd mechanically eﬃcient structural designs and shapes was initiated mainly in the ﬁelds of mechanical engineering and aeronautical engineering, which established the ﬁeld known as structural optimization. Many practically acceptable results have been developed for application to automobiles and aircraft. Some examples are structural components, including the wings of aircraft and engine mounts of automobiles, which can be fully optimized using eﬃcient shape optimization techniques. In contrast, regarding civil engineering and architectural engineering, structural optimization is diﬃcult to apply because structures in these ﬁelds are not mass products: structures are designed in accordance with their speciﬁc design requirements. Furthermore, the structure’s shape and geometry are determined by a designer or an architect in view of nonstructural performance, including the aesthetic perspective. Therefore, the main role of structural engineers is often limited to selection of materials, determination of member sizes through structural analyses, planning details of the construction process, and so on. However, for special structures, such as shells, membrane structures, spatial long-span frames, and highrise buildings, the structural shape should be determined in view of the responses against static and dynamic loads. In truth, the beauty of the structural form is related closely to the mechanical performance of the structure. Therefore, cooperation between designers and structural engineers is very important in designing such structures. Even for building frames, because of the recent trend of performance-based design, optimization has been identiﬁed as a powerful tool for designing structures under constraints imposed on practical performance measures, including elastic/plastic stresses and displacements under static/dynamic design loads. Furthermore, recent rapid advancements in the areas of computer hardware and software enabled us to carry out structural analysis many times to obtain optimal or approximately optimal designs. Optimization of real-world structures with realistic objective function and constraints is possible through quantitative evaluation of nonstructural performance criteria, e.g., aesthetic properties, and life-cycle costs, including costs of construction, fabrication, and maintenance. Many books describing structural optimization have been published since the 1960s; e.g., Hemp (1973), Rozvany (1976), Haug and Cea (1981), Haftka, G¨ urdal, and Kamat (1990), Papalambos and Wilde (2000), Bendsøe and Sigmund (2003), Arora (2004), etc. These books are mainly classiﬁable into the following three categories:

v

vi

Preface 1. Basic theories and methodologies for optimization with examples of small structural optimization problems. 2. Continuum-based approaches for application to mechanical and aeronautical structures. 3. Theoretical and analytical results of structural optimization in earlier times without the assistance of computer technology.

Using books of the ﬁrst category, readers can learn only the concepts and some diﬃcult theories of structural optimization without application to largescale structures. On the other hand, for the books of the second category, a good background in applied mathematics and continuum mechanics is needed to fully understand the basic concepts and methods. Unfortunately, most researchers, practicing engineers, and graduate students in the ﬁeld of civil engineering have no such background and are not strongly interested in the basic theories or methods of structural optimization. Also, in mechanical engineering, the ﬁnite element approach is used for practical applications, and complex practical design problems are solved in a ﬁnite dimensional formulation. The derivatives of objective and constraint functions, called design sensitivity coeﬃcients, should be computed if a gradient-based approach is used for structural optimization. However, most methods of design sensitivity analysis are developed mainly for a continuum utilizing variational principles, for which sensitivity coeﬃcients are to be computed for a functional, such as compliance that can be formulated in an integral of a bilinear form of response. For ﬁnite dimensional structures, including trusses and frames, variational formulations are not needed, and sensitivity coeﬃcients can be found simply by diﬀerentiating the governing equations in a matrix-vector form. Another important aspect of optimization in civil engineering is that the design variables often have discrete values: the frame members are usually selected from a pre-assigned list or catalog of available sections. Furthermore, some traditional layouts are often used for plane and spatial trusses and for latticed domes. Therefore, the optimization problem often turns out to be a combinatorial problem, a fact that is not fully introduced into most books addressing the study of structural optimization. This book introduces methodologies and applications that are closely related to design problems of ﬁnite dimensional structures, to serve thereby as a bridge between the communities of structural optimization in mechanical engineering and the researchers and engineers in civil engineering. The book provides readers with the basics of optimization of frame structures, such as trusses, building frames, and long-span structures, with descriptions of various applications to real-world problems. Recently, many eﬃcient techniques of optimization have been developed for convex programming problems, e.g., semideﬁnite programming and interior point algorithms, which are extensions of the approaches used for linear

Preface

vii

and quadratic programming problems. The book introduces application of these methods to optimization of ﬁnite-dimensional structures. Approximate methods resembling the conventional optimality criteria approaches have also been developed with no reference to the pioneering papers in the 1960s and 1970s. Therefore, it is extremely important to describe their development history to young researchers so that similar methods are not re-developed with no knowledge related to conventional approaches. For that reason, another purpose of this book is to present the historical development of the methodologies and theorems on optimization of frame structures. The book is organized as follows: In Chapter 1, the basic concepts and methodologies of optimization of trusses and frames are presented with illustrative examples. Traditional problems with constraints on limit loads, member stresses, compliance, and eigenvalues of vibration are described in detail. A brief introduction is also presented for multiobjective structural optimization, and the shape and topology optimization of trusses. In Chapter 2, the method of design sensitivity analysis, which is a necessary tool for optimization using nonlinear programming, is presented for various response quantities, including static response, eigenvalue of vibration, transient response for dynamic load, and so on. All formulations are written in matrix-vector form without resort to variational formulation to support ready comprehension by researchers and engineers. In Chapter 3, details of truss topology optimization are described, including historical developments and diﬃculties in problems with stress constraints and multiple eigenvalue constraints. Recently developed formulations by semidefinite programming and mixed integer programming are introduced. Applications to plane and spatial trusses are demonstrated. Chapter 4 presents methods for conﬁguration optimization for simultaneously optimizing the geometry and topology of trusses. Diﬃculties in optimization of regular trusses are extensively discussed, and an application to generating a link mechanism is presented. Chapter 5 summarizes various results of optimization of building frames. Uniqueness of the optimal solution of a regular frame is ﬁrst investigated, and applications of parametric programming are presented. Multiobjective optimization problems are also presented for application to seismic design, and a simple heuristic method based on local search is presented. In Chapter 6, as a unique aspect of this book, optimization results are presented for spatial trusses and latticed domes. Simple applications of nonlinear programming and heuristic methods are ﬁrst introduced, and the spatial variation of seismic excitation is addressed in the following sections. The trade-oﬀ designs between geometrical properties and stiﬀness under static loads are shown for arch-type frames and latticed domes described using parametric curve and surface. Mathematical preliminaries and basic methodologies are summarized in the Appendix, so that readers can understand the details, if necessary, without the

viii

Preface

exposition of tedious mathematics presented in the main chapters. Various methodologies speciﬁcally utilized in some of the sections, e.g., the response spectrum approach for seismic response analysis, are also explained in the Appendix. Also, many small examples that can be solved by hand or using a simple program are presented in the main chapters. Therefore, this book is self-contained, and easily used as a textbook or sub-textbook in a graduate course. The author would like to deliver his sincere appreciation to Prof. Tsuneyoshi Nakamura, Prof. Emeritus of Kyoto University, Japan, for supervising the author’s study for master’s degree and Ph.D. dissertation on structural optimization. Supervision by Prof. Jasbir S. Arora of The University of Iowa during the author’s sabbatical leave is also acknowledged. The numerical examples in this book are a compilation of the author’s work on structural optimization at Kyoto University, Japan, during the period 1985–2010. The author would like to extend his appreciation to researchers for collaborations on the studies that appear as valuable contents in this book, namely, Prof. Naoki Katoh of the Dept. of Architecture and Architectural Engineering, Kyoto University; Prof. Shinji Nishiwaki of the Dept. of Mechanical Engineering and Science, Kyoto University; Prof. Hiroshi Tagawa of the Dept. of Environmental Engineering and Architecture, Nagoya University; Prof. Yoshihiro Kanno of the Dept. of Mathematical Informatics, University of Tokyo; Prof. Peng Pan of the Dept. of Civil Engineering, Tsinghua University, P. R. China; Dr. Takao Hagish*ta of Mitsubishi Heavy Industries; Mr. Yuji Kato of JSOL Corporation; Mr. Takuya Kinosh*ta, Mr. Shinnosuke Fujita, and Mr. Ryo Watada, graduate students in the Dept. of Architecture and Architectural Engineering, Kyoto University. The author would also like to thank again Prof. Yoshihiro Kanno of University of Tokyo for checking the details of the manuscript. The assistance of Ms. Kari Budyk and Ms. Leong Li-Ming of CRC Press and Taylor & Francis in bringing the manuscript to its ﬁnal form is heartily acknowledged. January 2010

Makoto Ohsaki

Contents

Preface

v

1 Various Formulations of Structural Optimization 1.1 Overview of structural optimization . . . . . . . . . 1.2 History of structural optimization . . . . . . . . . . 1.3 Structural optimization problem . . . . . . . . . . . 1.3.1 Continuous problem . . . . . . . . . . . . . . 1.3.2 Discrete problem . . . . . . . . . . . . . . . . 1.4 Plastic design . . . . . . . . . . . . . . . . . . . . . 1.5 Stress constraints . . . . . . . . . . . . . . . . . . . 1.6 Fully-stressed design . . . . . . . . . . . . . . . . . . 1.6.1 Stress-ratio approach . . . . . . . . . . . . . . 1.6.2 Single loading condition . . . . . . . . . . . . 1.6.3 Multiple loading conditions . . . . . . . . . . 1.7 Optimality criteria approach . . . . . . . . . . . . . 1.8 Compliance constraint . . . . . . . . . . . . . . . . . 1.8.1 Problem formulation and sensitivity analysis 1.8.2 Optimality conditions . . . . . . . . . . . . . 1.8.3 Reformulation of the optimization problem . 1.8.4 Convexity of compliance . . . . . . . . . . . . 1.8.5 Other topics on compliance optimization . . . 1.9 Frequency constraints . . . . . . . . . . . . . . . . . 1.10 Conﬁguration optimization of trusses . . . . . . . . 1.11 Multiobjective structural optimization . . . . . . . . 1.11.1 Basic concepts . . . . . . . . . . . . . . . . . 1.11.2 Problem formulation . . . . . . . . . . . . . . 1.12 Heuristic approach . . . . . . . . . . . . . . . . . . . 1.13 Simultaneous analysis and design . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . . .

1 1 3 5 5 10 11 14 17 17 20 23 25 29 29 31 34 39 42 43 48 50 50 51 52 55

2 Design Sensitivity Analysis 2.1 Overview of design sensitivity analysis 2.2 Static responses . . . . . . . . . . . . 2.2.1 Direct diﬀerentiation method . 2.2.2 Adjoint variable method . . . . 2.3 Eigenvalues of free vibration . . . . . 2.3.1 Simple eigenvalue . . . . . . . . 2.3.2 Multiple eigenvalues . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

59 59 62 62 66 69 69 73

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

ix

x

Contents 2.4 2.5

Linear buckling load . . . . . . . . . Transient responses . . . . . . . . . 2.5.1 Direct diﬀerentiation method 2.5.2 Adjoint variable method . . . Nonlinear responses . . . . . . . . . Shape sensitivity analysis of trusses

. . . . . .

76 78 78 79 81 83

3 Topology Optimization of Trusses 3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2 Michell truss . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.3 Topology optimization problem . . . . . . . . . . . . . . . . . 3.4 Optimization methods . . . . . . . . . . . . . . . . . . . . . . 3.5 Stress constraints . . . . . . . . . . . . . . . . . . . . . . . . 3.5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . 3.5.2 Governing equations . . . . . . . . . . . . . . . . . . . 3.5.3 Discontinuity in stress constraint . . . . . . . . . . . . 3.5.4 Discontinuity due to member buckling . . . . . . . . . 3.5.5 Mathematical programming approach . . . . . . . . . 3.5.6 Problem with stress and local constraints . . . . . . . 3.6 Mixed integer programming for topology optimization with discrete variables . . . . . . . . . . . . . . . . . . . . . . . . . . 3.6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . 3.6.2 Compliance minimization problem . . . . . . . . . . . 3.6.3 Stress constraints . . . . . . . . . . . . . . . . . . . . . 3.6.4 Numerical examples . . . . . . . . . . . . . . . . . . . 3.7 Genetic algorithm for truss topology optimization . . . . . . 3.7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . 3.7.2 Optimization considering nodal cost . . . . . . . . . . 3.7.3 Topological bit and ﬁtness function . . . . . . . . . . . 3.7.4 Numerical examples . . . . . . . . . . . . . . . . . . . 3.8 Random search method using exact reanalysis . . . . . . . . 3.8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . 3.8.2 Exact reanalysis . . . . . . . . . . . . . . . . . . . . . 3.8.3 Random search for topology optimization of trusses . 3.9 Multiple eigenvalue constraints . . . . . . . . . . . . . . . . . 3.9.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . 3.9.2 Multiple eigenvalues in optimal topology . . . . . . . . 3.9.3 Semideﬁnite programming for topology optimization . 3.9.4 Linear buckling constraint . . . . . . . . . . . . . . . . 3.9.5 Numerical examples . . . . . . . . . . . . . . . . . . . 3.10 Application of data mining . . . . . . . . . . . . . . . . . . . 3.10.1 Frequent item set of decent solutions . . . . . . . . . . 3.10.2 Topology mining of ground structures . . . . . . . . .

85 85 87 88 90 93 93 94 95 98 101 106

2.6 2.7

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

113 113 114 115 119 122 122 123 123 126 128 128 128 133 136 136 138 140 142 144 149 149 153

Contents 4 Conﬁguration Optimization of Trusses 4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . 4.2 General formulation and methodologies of conﬁguration mization . . . . . . . . . . . . . . . . . . . . . . . . . . 4.3 Optimization of a regular grid truss . . . . . . . . . . . 4.3.1 Problem formulation . . . . . . . . . . . . . . . . 4.3.2 Numerical examples . . . . . . . . . . . . . . . . 4.4 Generation of a link mechanism . . . . . . . . . . . . . 4.4.1 Introduction . . . . . . . . . . . . . . . . . . . . 4.4.2 Mechanical model of a link mechanism . . . . . . 4.4.3 Problem formulation . . . . . . . . . . . . . . . . 4.4.4 Numerical examples . . . . . . . . . . . . . . . .

xi

. . . opti. . . . . . . . . . . . . . . . . . . . . . . . . . .

159 159 160 166 166 172 174 174 174 178 181

5 Optimization of Building Frames 185 5.1 Overview of optimization of building frames . . . . . . . . . 185 5.1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . 185 5.1.2 Problem formulation . . . . . . . . . . . . . . . . . . . 186 5.1.3 Continuum approach . . . . . . . . . . . . . . . . . . . 192 5.1.4 Semi-rigid connections and braces . . . . . . . . . . . 192 5.1.5 Formulation of cost function . . . . . . . . . . . . . . . 197 5.2 Local and global searches of approximate optimal designs . . 198 5.2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . 198 5.2.2 Optimization problem and optimality conditions . . . 200 5.2.3 Local search of approximate optimal solutions . . . . . 202 5.2.4 Global search of approximate optimal solutions . . . . 206 5.2.5 Numerical example of a regular plane frame . . . . . . 208 5.3 Parametric optimization of frames . . . . . . . . . . . . . . . 214 5.3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . 214 5.3.2 Two-level decomposition of frames . . . . . . . . . . . 216 5.3.3 General concept of decomposition to subsystems . . . 220 5.3.4 Parametric multidisciplinary optimization problem . . 222 5.3.5 Optimization of plane frames . . . . . . . . . . . . . . 224 5.3.6 Optimization of a three-dimensional frame . . . . . . . 228 5.4 Local search for multiobjective optimization of frames . . . . 234 5.4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . 234 5.4.2 Heuristic approaches to combinatorial multiobjective programming . . . . . . . . . . . . . . . . . . . . . . . . . 235 5.4.3 Local search for multiobjective structural optimization 240 5.4.4 Properties of Pareto optimal solutions . . . . . . . . . 242 5.4.5 Numerical examples . . . . . . . . . . . . . . . . . . . 243 5.5 Multiobjective seismic design of building frames . . . . . . . 250 5.5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . 250 5.5.2 Formulation of the multiobjective programming problem . . . . . . . . . . . . . . . . . . . . . . . . . . . . 252 5.5.3 Optimization method . . . . . . . . . . . . . . . . . . 253

xii

Contents 5.5.4

Numerical examples . . . . . . . . . . . . . . . . . . .

254

6 Optimization of Spatial Trusses and Frames 259 6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 259 6.2 Seismic optimization of spatial trusses . . . . . . . . . . . . . 261 6.2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . 261 6.2.2 Design sensitivity analysis . . . . . . . . . . . . . . . . 262 6.2.3 Optimization against seismic excitations . . . . . . . . 263 6.3 Heuristic approaches to optimization of a spatial frame . . . 266 6.4 Shape optimization considering the designer’s preference . . 271 6.4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . 271 6.4.2 Description of an arch-type frame using a B´ezier curve 273 6.4.3 Shape optimization incorporating the designer’s preference . . . . . . . . . . . . . . . . . . . . . . . . . . . . 275 6.4.4 Sensitivity analysis with respect to control points . . . 277 6.4.5 Numerical examples . . . . . . . . . . . . . . . . . . . 278 6.5 Shape optimization of a single-layer latticed shell . . . . . . 280 6.5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . 280 6.5.2 Description of a latticed shell and formulation of the optimization problem . . . . . . . . . . . . . . . . . . 281 6.5.3 Numerical examples . . . . . . . . . . . . . . . . . . . 284 6.6 Conﬁguration optimization of an arch-type truss with local geometrical constraints . . . . . . . . . . . . . . . . . . . . . . 288 6.6.1 Direct assignments of geometrical constraints . . . . . 288 6.6.2 Optimization using a genetic algorithm . . . . . . . . 291 6.6.3 Numerical examples . . . . . . . . . . . . . . . . . . . 294 6.7 Seismic design for spatially varying ground motions . . . . . 295 6.7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . 295 6.7.2 Response to spatially varying ground motions . . . . . 295 6.7.3 Problem formulation and design sensitivity analysis . 299 6.7.4 Postoptimal analysis . . . . . . . . . . . . . . . . . . . 301 6.7.5 Numerical examples . . . . . . . . . . . . . . . . . . . 302 6.8 Substructure approach to seismic optimization . . . . . . . . 305 6.8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . 305 6.8.2 Frequency domain analysis for a secondary structure . 306 6.8.3 Optimization problem . . . . . . . . . . . . . . . . . . 309 6.8.4 Numerical examples . . . . . . . . . . . . . . . . . . . 310 Appendix A.1 Mathematical preliminaries . . . . . . . . . . . . . . A.1.1 Positive deﬁnite matrix and convex functions A.1.2 Rayleigh’s principle . . . . . . . . . . . . . . A.1.3 Singular value decomposition . . . . . . . . . A.1.4 Directional derivative and subgradient . . . . A.2 Optimization methods . . . . . . . . . . . . . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

. . . . . .

315 315 315 316 318 319 319

Contents

A.3

A.4

A.5 A.6

A.7

A.8

A.2.1 Classiﬁcation of optimization problems . . . . . . A.2.2 Nonlinear programming . . . . . . . . . . . . . . A.2.3 Dual problem . . . . . . . . . . . . . . . . . . . . A.2.4 Semideﬁnite programming . . . . . . . . . . . . . A.2.5 Combinatorial problem . . . . . . . . . . . . . . Heuristics . . . . . . . . . . . . . . . . . . . . . . . . . . A.3.1 Introduction . . . . . . . . . . . . . . . . . . . . A.3.2 Single-point-search heuristics . . . . . . . . . . . Multiobjective programming . . . . . . . . . . . . . . . A.4.1 Deﬁnition of multiobjective programming . . . . A.4.2 Constraint approach . . . . . . . . . . . . . . . . A.4.3 Linear weighted sum approach . . . . . . . . . . A.4.4 Goal programming . . . . . . . . . . . . . . . . . Parametric structural optimization problem . . . . . . . Parametric curves and surfaces . . . . . . . . . . . . . . A.6.1 B´ezier curve . . . . . . . . . . . . . . . . . . . . . A.6.2 B´ezier surface . . . . . . . . . . . . . . . . . . . . A.6.3 Adjoint curve . . . . . . . . . . . . . . . . . . . . Response spectrum approach . . . . . . . . . . . . . . . A.7.1 SRSS method . . . . . . . . . . . . . . . . . . . . A.7.2 CQC method . . . . . . . . . . . . . . . . . . . . A.7.3 Design response spectrum . . . . . . . . . . . . . A.7.4 Sensitivity analysis of mean maximum response . List of available standard sections of beams and columns

xiii . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . . . . . . . . . . . . . . . . . . .

319 321 334 336 338 340 340 341 345 345 347 348 349 350 353 353 356 357 359 359 361 362 363 364

References

367

Index

407

Author Index

417

Chapter 1 Various Formulations of Structural Optimization

Various formulations of optimization of ﬁnite dimensional structures are presented in this chapter. The concepts of structural optimization are ﬁrst presented in Sec. 1.1 followed by historical review in Sec. 1.2. The basic formulations are presented in Sec. 1.3 with an illustrative example. The simple optimization approach to plastic design that is formulated as a linear programming problem is presented in Sec. 1.4. Optimization results under stress constraints are shown in Sec. 1.5. The approximate method called fully-stressed design (FSD) is presented in Sec. 1.6 with investigation of the relation between optimum design and FSD. The optimality criteria approach to a problem with displacement constraints is presented in Sec. 1.7. Problems concerning the compliance and frequency of free vibration as measures of static and dynamic stiﬀness are extensively studied in Secs. 1.8 and 1.9, respectively. An example of shape and topology optimization of a truss is presented in Sec. 1.10 as an introduction to Chaps. 3 and 4. The basic formulation of multiobjective structural optimization programming and various methodologies of heuristics are shown in Secs. 1.11 and 1.12, respectively, as an introduction to several sections in the following chapters. Finally, developments in simultaneous analysis and design are summarized in Sec. 1.13.

1.1

Overview of structural optimization

In the process of designing structures in various ﬁelds of engineering, the designers and engineers make their best decisions at every step in view of structural and non-structural aspects such as stiﬀness, strength, serviceability, constructability, and aesthetic property. In other words, they make their optimal decisions to realize their best designs; hence, the process of structural design may be regarded as an optimum design even though optimality is not explicitly pursued. Structural optimization is regarded as an application of optimization methods to structural design. The typical structural optimization problem is formally formulated to minimize an objective function representing the structural

1

2

Optimization of Finite Dimensional Structures Structural analysis evaluate responses Sensitivity analysis compute gradients of responses

Optimization algorithm function values gradients of functions

update variables

FIGURE 1.1: Relations among structural analysis, optimization algorithm, and design sensitivity analysis for optimization using a nonlinear programming approach.

cost under constraints on mechanical properties of the structure. The total structural weight or volume is usually used for representing the structural cost. Even for the case in which the structural weight is not strongly related to the cost, it is very important that a feasible solution satisfying all the design requirements can be automatically found through the process of optimization. The mechanical properties include nodal displacements, member stresses, eigenvalues of vibration, and linear buckling loads. The structural optimization problem can be alternatively formulated to maximize a mechanical property under constraint on the structural cost. Although there are many possible formulations for structural optimization, e.g., minimum weight design and maximum stiﬀness design, the term structural optimization or optimum design is usually used for representing all types of optimization problems corresponding to structural design. In this book, we consider ﬁnite dimensional structures, such as frames and trusses, which are mainly used in civil and architectural engineering. In the typical process of structural optimization of ﬁnite dimensional structures, the cross-sectional properties, nodal locations, and member locations are chosen as design variables. There are many methods for structural optimization that are classiﬁed into • Nonlinear programming based on the gradients (sensitivity coeﬃcients or derivatives) of the objective and constraint functions, which is the most popular and straightforward approach. • Heuristic approaches, including genetic algorithm and simulated annealing, that do not need gradient information. In a nonlinear programming approach, the design variables are updated in the direction deﬁned by the sensitivity coeﬃcients of the objective function and constraints. The relations among structural analysis, optimization algorithm, and design sensitivity analysis for optimization using a nonlinear programming approach are illustrated in Fig. 1.1, where the arrows represent the direction of data ﬂow; i.e., sensitivity analysis is carried out at each step of

Various Formulations of Structural Optimization

3

optimization to provide gradients of responses for the optimization algorithm, and structural analysis is needed for sensitivity analysis and function evaluation at an optimization step (see Chap. 2 and Appendix A.2.2 for details of sensitivity analysis and nonlinear programming, respectively). There are several approaches to the classiﬁcation of structural optimization problems. In the ﬁeld of continuum structural optimization, shape optimization means the optimization of boundary shape, whereas the addition and/or removal of holes are allowed in topology optimization (Bendsøe and Sigmund 2003). In this book, we present various methodologies and results for optimization of ﬁnite dimensional structures, including rigidly jointed frames and pin-jointed trusses. Since optimization of trusses and frames was developed gradually in 1960s and 1970s by academic groups in diﬀerent geographical locations, several diﬀerent terminologies, e.g., conﬁguration, geometry, and layout, were used for representing the similar processes of shape and topology optimization; see, e.g., Dobbs and Felton (1969), Svanberg (1981), Lin, Che, and Yu (1982), Imai and Schmit (1982), Zhou and Rozvany (1991), Twu and Choi (1992), Bendsøe, Ben-Tal, and Zowe (1994), Dems and Gatkowski (1995), Ohsaki (1997b), Bojczuk and Mr´oz (1999), Stadler (1999), Evgrafov (2006), and Achtziger (2007). On the other hand, optimization of cross-sectional areas of trusses was traditionally called optimum design, design optimization, or structural optimization (Hu and Shield 1961; Prager 1974a; Rozvany 1976). However, the term sizing optimization was often used recently to distinguish it from shape optimization (Grierson and Pak 1993; Lin, Che, and Yu 1982; Zou and Chan 2005; Schutte and Groenwold 2003), and structural optimization covers all areas related to optimization of structures. In this chapter, we present a historical review and various formulations of optimization of ﬁnite dimensional structures.

1.2

History of structural optimization

The origin of structural optimization is sometimes credited to Galileo Galilei (1638), who investigated the optimal shape of a beam subjected to a static load. However, his approach was rather intuitive, and he did not establish any theoretical foundation of structural optimization. The intrinsic properties of minimizing or maximizing functions or functionals in physical phenomena in nature were noticed from ancient times as various minimum/maximum principles. The theoretical basis of minimum principles as a foundation of modern optimization was investigated in the 18th century and established as the calculus of variation. The principle of minimum potential energy that leads to the shape of a hanging cable called catenary is extensively used nowadays for the design of ﬂexible structures, e.g., cable nets

4

Optimization of Finite Dimensional Structures

and membrane structures (Krishna 1979). The surface of the minimum area for the speciﬁed boundary shape in three-dimensional space is called minimal surface, which is equivalent to the surface with vanishing mean curvature, and can be achieved by a membrane with a uniform tension ﬁeld without external load or pressure. Therefore, the minimal surface is eﬀectively used as the ideal self-equilibrium shape for designing a membrane structure that does not have bending stiﬀness (Otto 1967, 1969). Papers by Michell (1904), Maxwell (1890), and Cilly (1900) are often cited as the ﬁrst paper that mentioned the basic idea of topology optimization; see Sec. 3.1 for the history of topology optimization. However, the so-called Michell truss or Michell structure has an inﬁnite number of members; hence, it did not lead to any practical development until the 1950s, when the properties of the optimal plastic design of frames were investigated (Foulkes 1954; Drucker and Shield 1961; Heyman 1959). We do not discuss the history of optimization of continuum structures such as plates and shells, because the scope of this book is limited to ﬁnite dimensional structures. A comprehensive literature review of early developments of structural optimization is found in Bradt (1986), which was originally published by the Polish Academy of Science, and includes about 300 entries up to the 1950s starting with the book by Galileo Galilei (1638), and more than 1800 entries for the period 1960–1980. In the 1950s, optimality conditions were studied for the plastic design of frames (Foulkes 1954; Drucker and Shield 1961). In the 1960s, conditions or criteria of optimality were derived utilizing minimum principles for several performance measures of structures (Sewell 1987). Hu and Shield (1961) investigated the uniqueness of optimal plastic design. Taylor (1967) derived the optimality condition for a vibrating rod with speciﬁed natural frequency using Hamilton’s principle or the principle of least action. Prager and Taylor (1968) developed optimality conditions for sandwich beams considering constraints on compliance, natural frequency, buckling load, and plastic limit load, using minimal total potential energy, Rayleigh’s principle, and lower- and upperbound theorems of limit analysis, respectively. Prager (1972, 1974a) summarized the optimality conditions corresponding to various types of constraints, including the case of multiple constraints. Plastic design of frames was extensively studied in the 1960s and 1970s, because analytical and/or computationally inexpensive methods can be used for this problem. Prager (1971) developed conditions for an optimal frame, subjected to alternative loads, exhibiting the so-called Foulkes mechanism. Adeli and Chyou (1987) presented a kinematic approach using automatic generation of independent mechanisms (see Hemp (1973) for various early developments in optimal plastic design). In the 1970s, when the computer power was still not strong enough to use mathematical programming approaches to optimization of real-world structures, optimality criteria (OC) approaches were widely used for ﬁnite dimensional structures. The modern discrete OC approaches to trusses and frames were initiated by Venkayya, Khot, and Berke (1973). Dobbs and Nelson (1975)

Various Formulations of Structural Optimization

5

developed the OC approach to truss design. Reviews of OC approaches are found in Berke and Venkayya (1974) and Venkayya (1978). Owing to the rapid development of computer hardware and software technologies, many numerical approaches were developed in the 1980s and 1990s to obtain optimization results for practical problems. Developments in this period can be found in many books, e.g., Arora (2007), Adeli (1994), Burns (2002), and Haftka, G¨ urdal, and Kamat (1990). It should be noted that the preferred terminologies for structural optimization vary with age. As noted earlier, structural optimization of trusses covered only optimization of cross-sectional properties in the 1950s and 1960s. However, sizing optimization was recently used to distinguish it from shape and topology optimization. Optimality conditions were ﬁrst called Kuhn-Tucker conditions; however, the name was corrected to Karush-Kuhn-Tucker conditions in the 1980s. Multiple load sets for formulation of constraints on static responses were called alternative loads until the 1970s; however, they are now usually called multiple loading conditions or multiple load sets. Furthermore, framed structure was used for representing ﬁnite dimensional structures, including pin-jointed trusses and rigidly jointed frames; however, they are classiﬁed into trusses and frames, respectively, in recent literature. In this book, we use up-to-date terminology, for consistency, even for describing the results of papers in the early stages of development.

1.3 1.3.1

Structural optimization problem Continuous problem

If the design variables can vary continuously, i.e., can have real values, and the objective and constraint functions are continuous and diﬀerentiable with respect to the variables, the structural optimization problem can be formulated as a nonlinear programming (NLP) problem. Let A = (A1 , . . . , Am )⊤ denote the vector of m design variables. For a sizing design optimization problem, A represents the cross-sectional areas of truss members, heights of the sections of frame members, etc. For a geometry optimization problem, A may represent the nodal coordinates of trusses and frames. All vectors are assumed to be column vectors throughout this book. The number of design variables is often reduced using the approach called design variable linking, utilizing, e.g., the symmetry properties of the structure. The requirements to be considered in practical applications can also be used for reducing the number of variables; e.g., the beams in the same story of a building frame should have the same section. However, in the following, we assume that each variable can vary independently, and, for trusses and frames, Ai belongs to member i, for simplicity.

6

Optimization of Finite Dimensional Structures

Consider an elastic ﬁnite dimensional structure subjected to static loads. The vector of state variables representing the nodal displacements is denoted by U = (U1 , . . . , Un )⊤ , where n is the number of degrees of freedom. In most of the design problems in various ﬁelds of engineering, the design requirements for responses such as stresses and displacements are given with inequality constraints speciﬁed by design codes: Hj (U(A), A) ≤ 0, (j = 1, . . . , nI )

(1.1)

where nI is the number of inequality constraints. Generally, there exist equality constraints on the response quantities; e.g., an eigenvalue of vibration should be exactly equal to the speciﬁed value. However, we consider inequality constraints only, for simple presentation of formulations. The constraint function Hj (U(A), A) depends on the design variables implicitly through the displacement (state variable) vector U(A) and also directly on the design variables. For example, the axial force Ni of the ith member of a truss is given using a constant n-vector bi , deﬁning the stressdisplacement relation as Ni = Ai b⊤ (1.2) i U(A) which depends explicitly on Ai and implicitly on A through U(A). L The upper and lower bounds, which are denoted by AU i and Ai , respectively, are usually given for the design variable Ai due to the restriction in fabrication and construction. The objective function, e.g., the total structural volume, is denoted by F (A). Then the structural optimization problem is formulated as Minimize

F (A)

(1.3a)

subject to Hj (U(A), A) ≤ 0, (j = 1, . . . , n ) I

ALi

≤ Ai ≤

AU i ,

(i = 1, . . . , m)

(1.3b) (1.3c)

Problem (1.3) is classiﬁed as an NLP problem, because U(A) is a nonlinear function of A; see Appendix A.2.2 for details of NLP. The constraints (1.3c) are called side constraints, bound constraints, or box constraints, which are treated separately from the general inequality constraints (1.3b) in most of the optimization algorithms. As is seen from the deﬁnition of constraints in (1.3b), the diﬀerential coefﬁcients of U(A) with respect to A, called design sensitivity coeﬃcients, are needed when solving Problem (1.3) using a gradient-based NLP algorithm. For convenience in deriving the conditions to be satisﬁed at the optimal solution, the constraint function with respect to A only is deﬁned as e j (A) = Hj (U(A), A) H

(1.4)

If the side constraints are treated separately from the general inequality constraints, the conditions for optimality are derived using the Lagrangian

Various Formulations of Structural Optimization

7

ψ(A, µ) deﬁned as n ∑ I

ψ(A, µ) = F (A) +

e j (A) µj H

(1.5)

j=1

where µ = (µ1 , . . . , µnI )⊤ (≥ 0) is the vector of Lagrange multipliers. The necessary conditions for local optimality, which are called KarushKuhn-Tucker conditions or simply KKT conditions, are given as ∂ψ ≥ 0 for Ai = ALi ∂A i ∂ψ (1.6) = 0 for ALi < Ai < AU i ∂A i ∂ψ ≤ 0 for Ai = AU i ∂Ai where n ∑ ej ∂ψ ∂F ∂H = + µj , (i = 1, . . . , m) ∂Ai ∂Ai j=1 ∂Ai

(1.7)

e j ≤ 0, µj ≥ 0, µj H e j = 0, (i = 1, . . . , nI ) H

(1.8)

I

The third equation in (1.8) is called complementarity conditions (see Appendix A.2.2.3 for details of the optimality conditions). Conditions (1.6)–(1.8) are the necessary and suﬃcient conditions for local optimality, if all the objective and constraint functions are locally convex. Furthermore, (1.6)–(1.8) are suﬃcient conditions for global optimality, if all the objective and constraint functions are globally convex. For problems with real variables and continuously diﬀerentiable functions, the optimal solutions are found using various approaches of mathematical programming. If the objective function and the constraints are linear functions of the design variables, the problem is formulated as a linear programming (LP) problem, and the optimal solutions are easily found using the standard approach called the simplex method (Luenberger 2003) or the relatively new approach called the interior-point method (Karmarkar 1984; Gondzio 1995). If the objective and the constraint functions are nonlinear, various approaches of the NLP problem can be used (Fiacco and Cormic 1968; Mangasarian 1969; Pierre and Lowe 1975; Peressini, Sullivan, and Uhl 1988; BenIsrael, Ben-Tal, and Zolbec 1981; Bersekas 1982). However, there is no approach that is applicable to any type of NLP problem; i.e., the most suitable method should be appropriately chosen for each problem at hand. Furthermore, the method should be selected with regard to the desired accuracy and computational cost for optimization. One of the most popular approaches is sequential quadratic programming (Gill, Murray, and Saunders 2002), which

8

Optimization of Finite Dimensional Structures 2P (3)

P 5

(4)

1

2 3

x

y

FIGURE 1.2:

H

4

(1)

(2)

W1

W2

W1

A ﬁve-bar plane truss.

is used for most of the examples of the application of NLP in this book. Readers may refer to Appendix A.2.2 for details of NLP. Example 1.1 As a simple example of structural optimization, consider a ﬁve-bar plane truss, as shown in Fig. 1.2, subjected to vertical static loads. The intersecting members 3 and 4 are not connected with each other. The ﬁve bars are classiﬁed into Groups 1 and 2, consisting of members {1, 2, 5} and {3, 4}, respectively. The members in Group i (i = 1, 2) have the same cross-sectional area Agi , and let Ag = (Ag1 , Ag2 )⊤ . The sum of the lengths of members in the ith group is denoted by Lgi . Then the total structural volume C(Ag ), which is taken as the objective function, is deﬁned as C(Ag ) = Ag1 Lg1 + Ag2 Lg2

(1.9)

For a simple illustration of the problem, the constraints are given, as follows, for the y-directional displacement U3 of node 3, which is assumed to be positive, and the stress σ4 of member 4, which is assumed to be negative: U3 ≤ U3U , σ4L ≤ σ4

(1.10)

where U3U and σ4L are the speciﬁed upper bound of U3 and the lower bound of e 1 (Ag ) σ4 , respectively. The constraints are formulated using the function H g of A only: e 1 (Ag ) = U3 (Ag ) − U3U ≤ 0, H e 2 (Ag ) = −σ4L − σ4 (Ag ) ≤ 0 H

(1.11) 2

Let W1 = W2 = H = 1 m in Fig. 1.2. The elastic modulus is 200 kN/mm , and P = 10.0 kN. The bounds for the displacement and stress are U3U = 2 1.25 mm and σ4L = −0.06 kN/mm . The set of solutions satisfying U3 = U3U and σ4 = σ4L is shown in the solid lines in Fig. 1.3 that are drawn in the design variable space. The gray area

Various Formulations of Structural Optimization

9

U 3 = U 3U 500 400 300

A2g

b 200

a

σ 4 = σ4L

100 0 0

100

200

300

400

500

A1g

FIGURE 1.3:

Feasible region and optimal solutions of the ﬁve-bar truss.

is the feasible region satisfying the two constraints with equality. From (1.9), we obtain 1 Lg (1.12) Ag2 = g C(Ag ) − 1g Ag1 L2 L2 The solution on each dotted line in Fig. 1.3 has the same values of C. Therefore, if Ag1 and Ag2 can take real values, the point ‘a’ with the coordinates (Ag1 , Ag2 ) = (184.33, 198.90) in the design variable space corresponds to the optimal solution. In order to verify the optimality of the solution, the sensitivity coeﬃcients are obtained at the optimal solution as ∂U3 (Ag ) ∂U3 (Ag ) = −0.013112, = −0.018015, g ∂A1 ∂Ag2 ∂σ4 (Ag ) ∂σ4 (Ag ) = 0.45650, = 0.20541 g ∂A1 ∂Ag2

(1.13)

The sensitivity coeﬃcients of the objective function are easily computed from the member lengths as ∂C ∂C = 3828.4, = 4472.1 ∂Ag1 ∂Ag2

(1.14)

Then, from the second equation in (1.6) with i = 1 and 2, while Ai is replaced by Agi , the positive Lagrange multipliers are found as λ1 = 1.8680 × 105 and λ2 = 2.2694 × 107 . Hence, the optimality conditions are satisﬁed at the solution (Ag1 , Ag2 ) = (184.33, 198.90). As is seen in the above example, the optimal solution can be found for a simple truss graphically in the design variable space, if we have only two design

10

Optimization of Finite Dimensional Structures

variables. However, for larger structures with more design variables, the optimal solutions are to be found numerically using a mathematical programming approach or a heuristic approach.

1.3.2

Discrete problem

Suppose a list or catalog of the available standard sections is given for a sizing optimization problem of a frame, and the list Ai of the cross-sectional properties of the ith member is given as Ai = {(A1i , Ii1 , Zi1 ), . . . , (Ari , Iir , Zir )}

(1.15)

where Aji is the cross-sectional area, Iij is the second moment of inertia, Zij is the section modulus of the jth candidate section for member i, and r is the number of available sections, which is the same for all members, for brevity. Note that other properties such as fully-plastic moment should be included if elastoplastic responses are to be considered; see Appendix A.8 for examples of section lists. Suppose that Ji = j (i = 1, . . . , m) indicates that the jth section in the list is assigned to the ith member, where m is the number of members. This way, the mechanical properties of the frame are deﬁned by the vector J = (J1 , . . . , Jm )⊤ of integer variables. Hence, the nodal displacement vector is a function of J that is denoted by U(J). The objective and the constraint functions are e j (J) = Hj (J, U(J)), also functions of J, which are written as F (J) and H respectively. Then the optimization problem with inequality constraints only is formulated as Minimize

F (J) e j (J) ≤ 0, (j = 1, . . . , nI ) subject to H Ji ∈ {1, . . . , r}, (i = 1, . . . , m)

(1.16a) (1.16b) (1.16c)

Since Problem (1.16) is an integer programming problem, which is equivalently called a combinatorial optimization problem, various methods, e.g., the branch-and-bound method and the branch-and-cut method, can be used (Horst and Tuy 1985; Horst, Pardalos, and Thoai 1995) (see Sec. 3.5 for application of the branch-and-bound method to topology optimization of trusses). For the example of the ﬁve-bar truss in Fig. 1.2, suppose Ag1 and Ag2 can take only integer values 100, 200, . . . . Then, the feasible designs satisfying (1.10) are plotted in the ﬁlled circle in Fig. 1.3, and the optimal solution exists at point ‘b’. Since the state variables are continuous functions of the design variables, a structural optimization problem turns out to be a mixed integer nonlinear programming (MINLP) problem (Floudas 1995) if the formulation of simultaneous analysis and design, see Sec. 1.13, is used considering the nodal displacements as independent variables. Arora (2002) classiﬁed the structural optimization problems into the following six categories:

Various Formulations of Structural Optimization

11

1. Continuous design variables; functions are twice continuously diﬀerentiable (standard NLP problem). 2. Mixed design variables; functions are twice continuously diﬀerentiable; discrete variables can have non-discrete values during the solution process (functions can be evaluated at non-discrete points). A conﬁguration optimization problem of a truss with discrete cross-sectional areas and continuous nodal coordinates belongs in this category. 3. Mixed design variables; functions are non-diﬀerentiable; discrete variables can have non-discrete values during the solution process. A conﬁguration optimization problem of a truss with discrete cross-sectional area, continuous nodal coordinates, and nodal cost deﬁned as a nondiﬀerential function of cross-sectional areas belongs in this category (see Sec. 4.3). 4. Mixed design variables; functions may or may not be diﬀerentiable; some of the discrete variables must have only discrete value in the solution process. A conﬁguration optimization problem with a list of candidate topologies and continuous nodal coordinates belongs in this category. 5. Mixed design variables; functions may or may not be diﬀerentiable; some of the discrete variables are linked to others; assignment of a value to one variable speciﬁes values for others. A frame optimization problem with discrete cross-sectional properties such as second moment of inertia linked with cross-sectional area belongs to this category. 6. Combinatorial problems; purely discrete non-diﬀerentiable problems. Optimization problems for selection of materials, location of supports, etc. belong to this category. Arora, Huang, and Hsieh (1994) summarized various methods of optimization with discrete variables.

1.4

Plastic design

Optimal plastic design is the simplest and classical problem of optimization of trusses and frames, which was extensively studied in the 1960s. Consider a truss consisting of a perfectly rigid-plastic material; i.e., the strain before yielding is negligibly small, and the stress after yielding is constant at the yield stress, which is assumed to be the same for all members. The truss is subjected to a vector of quasistatic proportional loads P = ΛP0 deﬁned by the load factor Λ and the constant load pattern vector P0 .

12

Optimization of Finite Dimensional Structures

The axial force vector is given as N = (N1 , . . . , Nm )⊤ , where m is the number of members. Let n denote the number of degrees of freedom. The equilibrium equations are formulated in terms of the n×m equilibrium matrix D as DN = ΛP0 (1.17) p ⊤ Let Np = (N1p , . . . , Nm ) denote the vector of tensile yield axial forces of the members. The yield axial force in compression is given for the ith member, ignoring member buckling, as −Nip . Then the yield condition is written as

−Nip ≤ Ni ≤ Nip , (i = 1, . . . , m)

(1.18)

Note that Nip is proportional to the cross-sectional area Ai as Nip = Ai σ p , where σ p is the tensile yield stress. First the plastic limit analysis problem is formulated as a linear programming (LP) problem. Utilizing the lower-bound theorem of plastic limit analysis (Shames and Cozzarelli 1997), we can obtain the plastic limit load factor through maximization of the load factor under constraints on the equilibrium equations and the yield conditions: Maximize Λ subject to − Np ≤ N ≤ Np DN = ΛP0

(1.19a) (1.19b) (1.19c)

which is an LP problem with variables Λ and N. Therefore, the plastic limit load can easily be obtained using a standard method of LP such as the simplex method. The problem of minimizing the total structural volume under constraint on limit load factor can also be formulated as an LP problem. Since Nip is proportional to Ai , the optimal design that minimizes the total structural volume can be obtained by minimizing Np⊤ L, where L = (L1 , . . . , Lm )⊤ is the vector of member lengths. The upper and lower bounds for Nip are denoted pU ⊤ ) and by NipU and NipL , respectively, with the vectors NpU = (N1pU , . . . , Nm pL pL pL ⊤ N = (N1 , . . . , Nm ) . The speciﬁed limit load factor is denoted by Λp . Then, the optimization problem is formulated as Minimize

Np⊤ L

(1.20a)

subject to − N ≤ N ≤ N p

p

p

DN = Λ P N

pL

≤N ≤N p

(1.20b) (1.20c)

pU

(1.20d)

where the variables are N and Np . Because Problem (1.20) is also an LP problem, this problem was extensively studied in the 1960s and is still important for application to the plastic design of trusses. Note that the plastic collapse mechanisms can be found as the Lagrange multipliers at the optimal

Various Formulations of Structural Optimization

13

3LP 0 2LP 0

M2

M4

(2)

M4

M2

M3

H (1) (3) M5

M1

W/2

FIGURE 1.4:

W/2

A simple plane frame.

solution of Problem (1.20), or by solving the dual of Problem (1.20) that is formulated on the basis of the upper-bound theorem of plastic limit analysis, which states that the smallest load factor corresponding to admissible strain and displacement rates deﬁnes the collapse load. The plastic design problem of a frame with concentrated plastic hinges can also be formulated as an LP problem, as follows, if the interaction between the axial force and bending moment on the yield condition is ignored (Adeli and Chyou 1987): Example 1.2 Consider a plane frame, as shown in Fig. 1.4, subjected to a proportional horizontal load 2ΛP 0 and a vertical load 3ΛP 0 simultaneously. The bending moments at the member ends and the center of the beam are denoted by M1 , . . . , M5 , as shown in Fig. 1.4, which illustrates the state where M1 , . . . , M5 are all positive. The numbers in parentheses are member numbers. The equilibrium equations are given as M2 + M1 M4 + M5 + = 2ΛP 0 , H H M2 − M3 M3 + M4 = 3ΛP 0 − W/2 W/2

(1.21)

The fully-plastic moment of member i is denoted by Mip . The yield conditions are then given as − M1p ≤ M1 ≤ M1p , −M1p ≤ M2 ≤ M1p , − M2p ≤ M2 ≤ M2p , −M2p ≤ M3 ≤ M2p , −M2p ≤ M4 ≤ M2p , − M3p ≤ M4 ≤ M3p , −M3p ≤ M5 ≤ M3p

(1.22)

14

Optimization of Finite Dimensional Structures

We assume that the tensile yield stress σ p and the compressive yield stress −σ p are the same, respectively, for all members, and the cross-section of each member is modeled as a sandwich section; i.e., the half of the crosssectional area Ai is concentrated at each ﬂange, and Mip is proportional to Ai as Mip = Ai ri σ p , where ri is the distance between the ﬂanges. Hence, the objective function that is proportional to the total structural volume of the frame is formulated as a function of Mp = (M1p , M2p , M3p )⊤ : F (Mp ) = M1p H + M2p W + M3p H

(1.23)

Since both the objective function and the constraints are linear functions of the variables M1 , . . . , M5 and Mp , the optimal solution can easily be found by solving an LP problem. For a simple case with P 0 = 1, Λp = 1, and H = W = 1, we obtain the optimal solution as M1p = 3/8 and M2p = M3p = 5/8 with M1 = M2 = 3/8, M3 = M4 = M5 = 5/8, and F (Mp ) = 13/8. It is well known that the Foulkes mechanism satisfying the following conditions exists at the optimal solution of a frame for the case in which Mip is proportional to Ai (Foulkes 1954): {

θi = λLi for Mip > MipL θi ≤ λLi for Mip = MipL

(1.24)

where θi is the sum of absolute values of the rotation rate of the plastic hinges in the ith member, and λ is a positive constant. Note that the upper bound for Mip is not considered, for brevity. Condition (1.24) suggests that the plastic energy dissipation rate per unit volume is the same for members with Mip > MipL . There have been many studies on plastic design since the 1960s (Tam and Jennings 1989). Multiple (alternative) loads are considered in some papers, e.g., Prager (1967) and Chan (1969). Munro and Chuang (1986) presented a fuzzy LP approach for the case in which uncertainty exists in the loads. A probabilistic LP approach to limit design under uncertainty was developed by Gavarini and Veneziano (1972).

1.5

Stress constraints

In view of structural design procedure in civil engineering based on allowable stress design criteria, it is very important to obtain an optimal design that satisﬁes stress and displacement constraints against design loads. In this section, we consider stress constraints only, for simple presentation of the

Various Formulations of Structural Optimization

15

optimization procedure. An approach to optimization of a truss under displacement constraints is demonstrated in Sec. 1.8. Another important aspect in structural design is that several loads, including static loads (self-weight, service load, snow load, etc.) and dynamic loads (wind load, seismic load, etc.), should be considered, and, in the practical design process, the dynamic loads are represented by equivalent static loads. Furthermore, the self-weight and service load are classiﬁed as long-term loads, while others are short-term loads. Therefore, diﬀerent bounds should be given for the stresses against each loading condition. Consider nP loading conditions (load patterns), and let the superscript k denote the variables and parameters corresponding to the kth loading condition. The upper and lower bounds for σik are denoted by σikU and σikL , respectively. Then the optimization problem for minimizing the total structural volume of a truss under stress constraints is formulated as Minimize subject to

m ∑

Ai Li i=1 σikL ≤ σik ≤ σikU , ALi ≤ Ai ≤ AU i ,

(1.25a) (i = 1, . . . , m; k = 1, . . . , nP ) (i = 1, . . . , m)

(1.25b) (1.25c)

where ALi and AU i are the lower and upper bounds for Ai , respectively. Note again that the nP load patterns are applied independently, and the stress constraints are assigned for each loading condition. Example 1.3 Optimum designs are found for a 10-bar truss, as shown in Fig. 1.5, subjected to vertical loads P1 and P2 , where the numbers with and without parentheses are node numbers and member numbers, respectively (Katoh, Ohsaki, and Tani 2002). Note that the intersecting members are not connected at their centers. A small lower bound ALi = 0.1 mm2 is given for Ai to prevent instability of the truss, while the upper bound is not given for Ai . The bounds 2 2 for the stresses are σikU = 0.2 N/mm and σikL = −0.2 N/mm . Optimal solutions are obtained using the optimization software package SNOPT Ver. 7.2 (Gill, Murray, and Saunders 2002), which utilizes the sequential quadratic programming; see Appendix A.2.2.5. First, consider a single loading condition (P1 , P2 ) = (0.0, 100.0 kN). The optimal cross-sectional areas and the optimal objective value are shown in the second column in Table 1.1. The optimal solution is also illustrated in Fig. 1.6, where the width of each member is proportional to its cross-sectional area. Note that Ai is equal to its lower bound in members 4, 5, 6, 8, and 10, which may be removed to obtain the statically determinate truss of optimal topology after ﬁxing the unstable node 4. The stress is equal to its upper or lower bound in each member with Ai > ALi . This process of topology optimization is called the ground structure approach; it is extensively studied in Chap. 3.

Optimization of Finite Dimensional Structures (1)

(3)

1

(5)

6

5 3 y

4 2

9 7

(2)

(4)

(6)

x

P1

P2

2.0 m

FIGURE 1.5:

10 8

2.0 m

16

2.0 m

A 10-bar truss.

TABLE 1.1: Optimal cross-sectional areas and structural volume of the 10-bar truss under stress constraints. Member number 1 2 3 4 5 6 7 8 9 10 Total volume (mm3 )

Ai (mm2 ) Single loading Multiple loading 999.931 825.107 500.069 674.893 707.010 459.771 0.100 421.531 0.100 211.499 0.100 0.100 499.937 499.909 0.100 0.129 707.017 706.978 0.100 0.100 8.00051 × 106 8.91591 × 106

Next, we obtain the optimal solution under stress constraints against multiple loading conditions (P1 , P2 ) = (0, 100.0 kN) and (100.0 kN, 0). The optimal cross-sectional areas, which are also illustrated in Fig. 1.7, and the objective value are listed in the last column of Table 1.1. The optimal objective value is 8.91591 × 106 mm3 , which is larger than that for the single loading condition. Only the members 6 and 10 connected to node 5 satisfy Ai = ALi , and node 5 cannot be removed, because the cross-sectional area of member 8 is larger than its lower bound. Note that a very strict tolerance of 10−10 is assigned for the constraints and optimality conditions in SNOPT. In fact, we can conﬁrm that the stresses for the ﬁrst load (P1 , P2 ) = (0, 100.0 kN) are 1 σ61 = σ10 = 0.18265 and σ81 = −0.2; the member 8 is fully stressed. Since the √ equilibrium condition N8 = − 2(N6 + N10 ) should be satisﬁed for member forces Ni of members 6, 8, and 10, the cross-sectional area should be larger

Various Formulations of Structural Optimization

17

P2

FIGURE 1.6: Optimal design of the 10-bar truss under single loading condition (P1 , P2 ) = (0.0, 100.0 kN).

P1

P2

FIGURE 1.7: Optimal solution for multiple loading conditions (P1 , P2 ) = (0, 100.0 kN) and (100.0 kN, 0).

than its lower bound if the stresses of members 6 and 10 are close to their upper or lower bounds. Hence, it should be noted that the member with Ai = ALi sometimes cannot be removed in the conventional ground structure approach of topology optimization, and the optimal topology may consist of many members with small cross-sectional areas. The result in Fig. 1.7 demonstrates that the optimal truss under multiple loading conditions is statically indeterminate even when the removal of members 6, 8, and 10 is allowed. It has been conﬁrmed that the stress is equal to its upper or lower bound against at least one loading condition for a member with Ai > ALi .

1.6 1.6.1

Fully-stressed design Stress-ratio approach

Consider again a simple optimization problem of a truss with stress constraints, and suppose only the lower bound ALi is given for the cross-sectional area Ai of the ith member. In a practical design process, obtaining a feasible

18

Optimization of Finite Dimensional Structures Total structural volume

1.1

1.05

1

0.95 0

10

20

30

40

50

Iterative step

FIGURE 1.8: Convergence history of the total structural volume of FSD of the 10-bar truss; solid line: r = 1, dashed line: r = 1.5.

solution is sometimes more important than minimizing an objective function. Furthermore, the results of the example in the previous section suggest that an optimal design can be obtained by ﬁnding the cross-sectional areas so that the stress of a member with ALi < Ai is equal to its upper or lower bound for at least one of the nP loading conditions. The design satisfying this condition is called a fully-stressed design (FSD). Note that the inequality constraints σikL ≤ σik ≤ σikU (k = 1, . . . , nP ) are to be satisﬁed by the member with Ai = ALi . In the fully-stressed design approach, a design satisfying these conditions is obtained by iteratively modifying the design variables. For a simple case of a single loading condition with σi1L = −σi1U for all members, the FSD can be obtained by the following simple iterative algorithm for updating the cross-sectional areas: ( (k+1)

Ai (k)

(k)

= Ai

|σi1 | σi1U

)r , (i = 1, . . . , m)

(1.26)

where Ai is the value of Ai at the kth step of iteration, and r is the parameter for controlling the convergence property, which is usually between 1 and 2. (k+1) Note that Ai is replaced by ALi if Ai < ALi is satisﬁed as the result of application of (1.26). The design update rule (1.26) called the stress-ratio approach, assumes that the modiﬁcation of the cross-sectional area of a member does not have any strong eﬀect on the axial forces of the members. For example, the axial forces of a statically determinate truss are determined only from the equilibrium equations and are independent of the cross-sectional areas. In this case, the stress of a member against the speciﬁed set of loads is inversely proportional to its cross-sectional area, and the FSD can be found within only one step of application of (1.26) with r = 1.

Various Formulations of Structural Optimization

19

The stress-ratio approach can also be eﬀectively used for building frames, for which constraints are given for the stress at each edge of the section at the member ends due to the bending moment and axial force. It is very convenient for investigating the nearly optimal load paths of a plane regular frame from the loaded nodes to the supports. If the variable Ai deﬁnes the size of the section with the dimension of length, e.g., height and width of the wide-ﬂange section, the appropriate value of the parameter r in (1.26) ranges between 1/3 and 1/2 (Mueller, Liu, and Burns 2002). Example 1.4 An FSD is found for the 10-bar truss in Fig. 1.5 under single loading condition (P1 , P2 ) = (0.0, 100.0 kN). A small lower bound, 0.1 mm2 , which is the same as the value in Example 1.3 in Sec. 1.5, is given for Ai to prevent instability of the truss and to compare the results. Fig. 1.8 shows the convergence history of the total structural volume divided by the value 8.91591 × 106 in Table 1.1 for the optimal solution under stress constraints subjected to the same single loading condition. The histories of A1 and A6 are also plotted in Figs. 1.9 (a) and (b), respectively, where only a small range is plotted for A6 , because A6 is equal to the lower bound, 0.1 mm2 , at the converged FSD. As is seen from these ﬁgures, the total structural volume and the cross-sectional areas converge monotonically to the optimal values within 30 steps, if r = 1. The convergence property is improved if a larger value, 1.5, is assigned for r; i.e., an approximate optimal solution can be found within 20 steps; however, some oscillation is observed at the early stage of iteration. Note that the total structural volume converges to the optimal value under stress constraints, and the cross-sectional areas of the FSD are the same as the optimal values in Table 1.1. This way, the optimal truss under stress constraints for a single loading condition can easily be found by the simple stress-ratio approach (1.26) if the absolute values of the upper- and lower-bound stresses are the same.

The relation between the FSD and optimum design under stress constraints has been extensively studied since the 1960s (Razani 1965; Kicher 1966; Patnaik and Dayaratnam 1970; McNeil 1971; Chern and Prager 1972; Nagtegaal 1973; Gunnlaugsson and Martin 1973) and was revisited mainly in the community of applied mathematics in the 1990s (Bendsøe and Sigmund 2003). However, it seems that the FSDs are not clearly deﬁned each case with ALi > 0 and ALi = 0. Here we do not assign an upper bound for the cross-sectional area, and deﬁne the FSD as follows (Nagtegaal 1973): • If ALi > 0, then the stress σik of a member with Ai > ALi should be equal to σikU or σikL for at least one loading condition; whereas σikL ≤ σik ≤ σikU should be satisﬁed by a member with Ai = ALi .

Optimization of Finite Dimensional Structures Cross-sectional area of member 1

20

1100 1000 900 800 700 600 500 400 0

10

20

30

40

50

Iterative step

Cross-sectional area of member 6

(a) member 1 1 0.8 0.6 0.4 0.2 0 20

30

40

50

60

70

80

90

100

Iterative step

(b) member 6 FIGURE 1.9: Convergence history of cross-sectional areas of FSD of the 10-bar truss; solid line: r = 1, dashed line: r = 1.5. • If ALi = 0, then the stress σik of a member with Ai > 0 should be equal to σikU or σikL for at least one loading condition; whereas no constraint exists on the stress of a nonexistent member with Ai = 0. Therefore, there is a discontinuity in the FSDs between the cases with ALi = 0 and ALi = e, where e is a small positive value. Note that the case with ALi = 0 corresponds to the topology optimization problem that is extensively investigated in Sec. 3.5.3.

1.6.2

Single loading condition

First we consider a truss subjected to a single loading condition, and let n denote the number of degrees of freedom. Suppose the truss consisting of m members is statically indeterminate; i.e., n < m with n being the number of degrees of freedom. The vectors of nodal displacements U1 = (U11 , . . . , Un1 )⊤

Various Formulations of Structural Optimization

1

4

2 3

21

x 5

y

P1 P1

FIGURE 1.10:

A statically indeterminate ﬁve-bar truss.

and member strains ε1 = (ε11 , . . . , ε1m )⊤ should satisfy the compatibility conditions ε1 = CU1 (1.27) where C is an m × n matrix that is deﬁned by the kinematic relations only. Suppose the truss is stable and C is full-rank; i.e., the rank of C is equal to n because n < m. Hence, we can eliminate U1 using (1.27) to express m − n components of ε1 with respect to the remaining n components. Therefore, using the constitutive relation σi1 = Eε1i with the elastic modulus 1 . Hence, the E, the m − n equations are obtained for the stresses σ11 , . . . , σm stress can be independently assigned only for n members, and the stresses of all the members of a statically indeterminate truss cannot generally be equal to the upper or lower bound. Consequently, for a truss to be fullystressed, the cross-sectional areas of at least m − n members should be equal to their lower bounds, and the stresses of the remaining n members should be appropriately assigned so that σi1 = σi1L or σi1 = σi1U for those members, and σi1L ≤ σi1 ≤ σi1U for the remaining m − n members with Ai = ALi . Example 1.5 Consider a symmetric ﬁve-bar truss, as shown in Fig. 1.10, where n = 2 and m = 5, i.e., the truss is statically indeterminate, and we can assign two member stresses independently. In the following examples, the units of load and length are omitted for brevity. The truss is symmetric with respect to the y-axis. The angles of members 1 and 2 from the x-axis are π/4 and π/3, respectively. If σ11 and σ31 are chosen as the independent variables, the stresses of the remaining members are given as √ √ 3 1 3− 3 1 1 σ2 = σ + σ , 2√ 1 4 √ 3 (1.28) 3 1 3+ 3 1 σ1 + σ3 , σ41 = − 2 4 σ51 = σ31 − σ11 which show that the stresses cannot have the same absolute value for all members. Therefore, generally Ai = ALi should be satisﬁed for three members

22

Optimization of Finite Dimensional Structures P

FIGURE 1.11:

1

2

L1

L2

A statically indeterminate two-bar truss.

so that the truss is fully stressed. Hence, the FSD is statically determinate if ALi = 0 for all members. Note that the stress of a nonexistent member can be computed from the strain because the two nodes (supports) connected to any member of this truss exist (see Sec. 3.5.3 for more details). For example, the total structural volume is minimized under conditions P1 = 1, P2 = 0, and ALi√= 0 for all members. If the bounds of stresses are given as σiU = −σiL = 1/ 2 for all members, the optimal cross-sectional areas are obtained as A1 = A√ A4 = 0. Then the stresses 5 = 1 and A2 = A3 =√ √ 2. Accordingly, 6/4, σ3 = 0, σ4 = − 6/4, are obtained as σ = 1/ σ = 1 2 √ σ5 = −1/ 2, and the truss is statically determinate and fully stressed. For this example, the truss is fully stressed even if a very small positive value e is assigned for ALi of all members because σiL ≤ σi ≤ σiU is satisﬁed by the nonexistent member 3.

Example 1.6 As another illustrative example, consider a statically indeterminate truss, as shown in Fig. 1.11, that has two colinearly located members, and assume P > 0. The bounds for the stress are given as σi1L = −σi1U , where σ11U and σ21U are not necessarily the same. The lower bounds for Ai are given as AL1 = AL2 = e, where e has a suﬃciently small positive value. If L2 = 2L1 and σ11U = σ21U = σ for a speciﬁed positive value σ, then the optimization for minimizing the total structural volume V leads to A1 ≃ P/σ and A2 = e, because member 2 is longer than member 1, and σ11 = σ and σ21 = −σ/2 are satisﬁed from the compatibility condition. Hence, the optimal solution is fully stressed. By contrast, if σ11U = σ and σ21U = 4σ, then the optimization leads to A1 = e and A2 ≃ P/(4σ), because the larger length of member 2 is compensated by the larger absolute value of the allowable stress; consequently, V is approximately equal to P L/(2σ). In fact, if we assume A2 = e, then A1 ≃ P/σ and, accordingly, V is approximately equal to P L/σ, which is larger than P L/(2σ). At the optimal solution with A2 ≃ P/(4σ), σ11 = 8σ, and σ21 = −4σ are satisﬁed. Hence, the optimal solution is not fully stressed; i.e., the stress constraint is violated by member 1. Therefore, the optimal solution may not be fully stressed if ALi > 0 and the stress bounds are not the same for all the members of a statically indeterminate truss. However, if ALi = 0, then the

Various Formulations of Structural Optimization

23

2 1

π /4 π /4

3 P1

P2

FIGURE 1.12:

A three-bar truss (Type 1).

optimal solution is (A1 , A2 ) = (0, P/(4σ)), which is fully stressed, because the stress constraint need not be satisﬁed by the nonexistent member 1.

1.6.3

Multiple loading conditions

Next we consider the truss under multiple loading conditions. The optimization problem under stress constraints is formulated as (1.25) without upper-bound cross-sectional area. Let nA denote the number of members for which the stress is equal to its lower or upper bound for at least one loading condition. If the truss is statically determinate, the axial force is independent of the cross-sectional areas, and nA = m should be satisﬁed; i.e., the truss is fully stressed. For a statically indeterminate truss, we can specify stresses for at most n × nP members, because the stresses of n members can be speciﬁed for each loading condition, as demonstrated in Example 1.5. Therefore, the stress can be equal to its lower or upper bound for all members if nP ≥ m/n (Patnaik and Dayaratnam 1970). However, it is well known that the optimal solution under multiple loading conditions is not generally fully stressed even for the case where σi1L = −σi1U and σi1U is the same for all members; i.e., σi1L ≤ σi1 ≤ σi1U may be satisﬁed by a member with Ai > 0 (Kicher 1966; Patnaik and Dayaratnam 1970; McNeil 1971; Gunnlaugsson and Martin 1973; Patnaik and Hopkins 1998). For a topology optimization problem with ALi = 0 for all members, the stress constraints may be violated by the nonexistent members, and the optimal truss may be statically determinate even for the multiple loading conditions; see Sec. 3.5 for details. Example 1.7 Consider a three-bar truss (Type 1), as shown in Fig. 1.12, which is subjected to two independent loads, P1 and P2 , respectively. From nP = 2, m = 3, and n = 2, we can see nP ≥ m/n is satisﬁed. Suppose √ √ P1 = P2 = 10, and the lengths of members 1, 2, and 3 are 2, 1, and 2, respectively. The

24

Optimization of Finite Dimensional Structures x

(1)

y

(2) (3)

2

1

3 (4)

Px Py

FIGURE 1.13:

A three-bar truss (Type 2).

lower-bound cross-sectional areas are given as ALi = 0.1 for all members; i.e., removal of a member is not allowed. L U √The optimal √ solution for σi = −σi = 10 for all members is (A1 , A2 , A3 ) = ( 2/2, 1/2, 2/2), where the total structural volume is 2.5. Note that the value of ALi does not have any eﬀect on the optimal solution if it is positive and not more than 1/2. The stresses are computed as (σ11 , σ21 , σ31 ) = (10, 0, −10) and (σ12 , σ22 , σ32 ) = (5, 10, 5). Therefore, the optimal truss is fully stressed. However, if ALi = 0, then a member may be removed to obtain a statically determinate optimal truss, and the stress constraint may be violated by a nonexistent member. √ In fact, √ if we remove member 2, then the optimal solution is (A1 , A3 ) = ( 2/2, 2/2), where the total structural volume is 2.0, which is less than 2.5 for the three-bar truss. If we assume that the elastic modulus E is equal to 1, then the vertical displacement of node 2 is 20. Therefore, the strain of member 2, although it does not exist, is 20, which is double the upper-bound stress.

Example 1.8 Next, consider a three-bar truss (Type 2), as shown in Fig. 1.13, where the coordinates of nodes and supports 1–4 are (0,0), (1000,0), (2000,500), and (500,1000), respectively. Three independent load sets (Px , Py ) = (5, 10), (−5, 10), and (−20, 10) are applied. The upper-bound stress is 0.2 and ALi = 0 for all members. The optimal cross-sectional areas and the maximum absolute value of stress of each member under three loading conditions are listed in the second column of Table 1.2. As is seen, the optimal truss is statically indeterminate. Because the absolute value of the stress of member 2 is less than 0.2 for any loading condition, the optimal truss is not fully stressed. In fact, the optimal solutions of two-bar trusses after removal of members 1, 2, and 3, respectively, have larger objective values, as shown in the third, fourth, and ﬁfth columns of Table 1.2, than that of the three-bar truss in the second column.

Various Formulations of Structural Optimization

25

TABLE 1.2:

Optimal cross-sectional areas, total structural volume, and maximum absolute values of stresses of the three-bar truss (Type 2) and corresponding two-bar trusses. three-bar two-bar 2, 3 1, 3 1, 2 A1 42.717 0 63.888 83.853 A2 36.063 89.443 0 139.75 A3 84.219 94.868 112.94 0 max |σ1k | 0.20000 0.60000 0.20000 0.20000 max |σ2k | 0.19709 0.20000 0.26429 0.20000 max |σ3k | 0.20000 0.20000 0.20000 0.30000 Total volume 2.3539 2.5000 2.5000 2.5000

1.7

Optimality criteria approach

An optimization method that directly solves the optimality criteria (optimality conditions) is called the optimality criteria approach (OC approach) (Venkayya, Khot, and Berke 1973; Berke and Venkayya 1974; Dobbs and Nelson 1975; Khot, Berke, and Venkayya 1978). This approach is very eﬀective for the case where the optimality conditions are written in a simple manner with explicit expressions of sensitivity coeﬃcients with respect to the design variables and the state variables. Furthermore, this approach is more eﬃcient in view of computational time and required memory than the gradient-based nonlinear programming (NLP) approaches; see Sec. 2.2 for sensitivity analysis of static responses, and Appendix A.2.2 for details of optimality conditions for general NLP problems. Another advantage of the OC approach is that the computer program is very simple. Therefore, in the 1960s and 1970s, when computer power was not suﬃcient for computing sensitivity coeﬃcients of the responses of moderately large structures many times for optimization, various studies on theoretical and computational aspects of the OC approaches were presented. For problems with general equality and inequality constraints, the OC approach is classiﬁed as a dual approach of NLP (Fleury 1979, 1980). Since the purpose here is to ﬁnd a solution that satisﬁes the constraints and optimality conditions, it is possible to use a Newton-Raphson iteration for solving these nonlinear equations (Khot, Berke, and Venkayya 1978). However, the recursive formulas, as presented below, are generally used in an OC approach. For the problem under stress constraints only, an approximate optimal solution can be easily found using the stress-ratio approach of fully-stressed design, as discussed in the previous section. Therefore, in this section, we consider an optimization problem of a truss under displacement constraints.

26

Optimization of Finite Dimensional Structures

Suppose, for simplicity, a constraint is given only for the jth displacement component as Uj ≤ UjU (1.29) where Uj is assumed to be positive. The objective function is the total structural volume, which is a monotonically increasing function of the crosssectional areas A = (A1 , . . . , Am )⊤ . On the other hand, the nodal displacements are generally decreasing functions of A, if the loads do not depend on A. Therefore, the displacement constraint (1.29) is considered to be active, i.e., satisﬁed with equality, at the optimal solution assuming that the lower bounds ALi for the cross-sectional areas are suﬃciently small. The upper bounds AU i are assumed to be suﬃciently large to ensure the existence of a feasible solution. In this case, the following condition is obtained from the optimality condition (1.6) for members with ALi < Ai < AU i : Li + µj

∂Uj =0 ∂Ai

(1.30)

where µj (≥ 0) is the Lagrange multiplier for the constraint (1.29). The term ∂Uj /∂Ai in (1.30) is the sensitivity coeﬃcient of Uj with respect to Ai that can be obtained eﬃciently using the adjoint variable method described in Sec. 2.2.2, because we have only one displacement component to be constrained. Let n denote the number of degrees of freedom. The axial force and the n × n stiﬀness matrix with respect to the global coordinates of the ith member are denoted by Ni and Ki , respectively. The displacement vector against the speciﬁed loads is denoted by U. The values corresponding to the virtual unit load at the jth displacement component are indicated by the superscript ( · )j . Then the following relation is derived from the adjoint variable method of design sensitivity analysis of static response: ∂Uj ∂Ki = −Uj⊤ U ∂Ai ∂Ai Li = − 2 Nij Ni Ai E

(1.31)

where E is the elastic modulus. Deﬁne Zi as Zi = µj

Nij Ni A2i E

(1.32)

Then the optimality condition (1.30) is written as Zi = 1

(1.33)

Let the superscript ( · )(k) denote a value at the kth step of iteration. For a statically determinate truss, Nij and Ni are independent of Ai . Therefore,

Various Formulations of Structural Optimization

27

assuming that µj is constant, the cross-sectional areas can be updated from (1.32) as follows: (k+1)

(k)

Ai

1

(k)

= (Zi ) 2 Ai

(1.34)

For a statically indeterminate truss, the cross-sectional area is updated by (k+1)

Ai

(k)

(k)

= (Zi )r Ai

(1.35)

where r is a parameter between 0 and 1 for controlling the convergence propL erty. Note that Ai is replaced with ALi and AU i , respectively, if Ai < Ai or Ai > AU after application of (1.35). i Next, we derive the update rule of the Lagrange multiplier. From (1.32) and (1.33), we obtain √ N j Ni (1.36) Ai = µj i E By using the principle of virtual unit load and the active constraint Uj = UjU , we have m ∑ Li UjU = Nij Ni (1.37) A E i i=1 Through incorporation of Ai in (1.36) into (1.37), the Lagrange multiplier µj is updated as √ 2 j m L ∑ N N i i i (k+1) √ (1.38) µj = U E U j i=1 Then we move to the next step of iteration. This way, the solution satisfying the constraint and optimality conditions is found by an iterative approach. Alternatively, a linear approximation can be used for recursively updating (k) Ai and µj (Khot, Berke, and Venkayya 1978). Multiplying (1 − α)Ai on both sides of (1.33) with a parameter 0 < α < 1, letting (k)

(1 − α)Ai

(k+1)

= Ai

(k)

− αAi

(1.39)

and rearranging the equation, we have the following update rule for Ai : (k+1)

Ai

(k)

(k)

= Ai [α + (1 − α)Zi ]

(1.40)

Linear approximation of Uj leads to the following requirement for the displacement constraint at the (k + 1)st step: (k)

Uj

+

m ∑ ∂Uj i=1

∂Ai

(k+1)

(Ai

(k)

− Ai ) = UjU

(1.41)

28

Optimization of Finite Dimensional Structures

Then, from (1.31), (1.37), (1.40), and (1.41), we obtain the following recursive formula for µj : µj

m ∑ Li (N j Ni )2 i=1

i E 3 A3i

(k)

(2 − α)(Uj − UjU ) = 1−α

(1.42)

The OC approach assumes that the member forces against the applied loads and the virtual unit load are insensitive to variation of the design variables, which is the same as the assumption for the stress-ratio approach (1.26) for fully-stressed design. This is better achieved if the responses are approximated with respect to the reciprocals of the cross-sectional areas (Schmit and Farshi 1974; Zhou and Haftka 1995). Let ai = 1/Ai , and regard Ai as a function of ai . Then we have 1 ∂Ai (1.43) =− 2 ∂ai ai and the sensitivity coeﬃcient of Uj with respect to ai is obtained from (1.31), (1.43), and Ai = 1/ai as ∂Uj ∂Uj ∂Ai = ∂ai ∂Ai ∂ai Li = − Nij Ni E

(1.44)

which does not explicitly depend on ai . Although the sensitivity of the total structural volume turns out to be dependent on ai , convergence of the recursive formulation is improved by using ai as a design variable. Because the number of analyses for computing the displacements against virtual unit load is proportional to the number of active or nearly active displacement constraints, this approach is eﬀective for the case with a small number of active displacement constraints. For extension of the OC approach to problems with stress constraints, a pair of unit self-equilibrium forces is applied at the two ends of each member with an active stress constraint. However, the OC approach can be successfully combined with the fully-stressed design approach for problems with stress and displacement constraints. Pereyra, Lawver, and Isenberg (2003) used an OC approach with a penalty function to optimize a building frame. For continuum structures such as beams, plates, and shells, a continuum-type optimality criteria (COC) approach was developed in the 1960s (Prager and Taylor 1968; Olhoﬀ and Taylor 1979; Rozvany 1989). Furthermore, OC and COC were combined to a discretized form of COC approach termed DCOC (Zhou and Rozvany 1992, 1993; Rozvany and Zhou 1994). Since the OC approach is simple and easy to implement, it is widely applied in many areas of heuristics, e.g., evolutionary structural optimization (ESO) (Yang, Xie, Steven, and Querin 1999a; Xie and Steven 1993) and cellular automaton (Canyurt and Hajela 2005).

Various Formulations of Structural Optimization

1.8 1.8.1

29

Compliance constraint Problem formulation and sensitivity analysis

The compliance is equivalent to the external work by static loads, which is also equivalent to twice the strain energy if the linear elastic response is considered. For the case without forced displacement, a smaller compliance leads to a stiﬀer structure against the speciﬁed loads. Consider a truss with m members and n degrees of freedom. Let K(A) denote the n × n stiﬀness matrix, which is a function of the vector A = (A1 , . . . , Am )⊤ of the cross-sectional areas. The displacement vector against the load vector P, which is independent of A, is denoted by U(A). The stiﬀness (equilibrium) equation for computing U(A) is written as K(A)U(A) = P

(1.45)

Using (1.45), the compliance W (U(A), A) is written as W (U(A), A) = U⊤ P = U⊤ KU ) ( 1 = 2 U⊤ P − U⊤ KU 2

(1.46)

i.e., the compliance is also equivalent to the total potential energy multiplied by −2. Note that the second argument A in W (U(A), A) indicates the explicit dependence on A through K(A). By diﬀerentiating W (U(A), A) in the last expression in (1.46), and using (1.45), we ﬁnd that the partial derivative of the compliance with respect to U vanishes as ∂W = 2(P − KU) = 0 ∂U

(1.47)

Because U is an implicit function of A, the compliance is written as a function of A only as f (A) = W (U(A), A) W (1.48) f (A). Then the optimization Let W U denote the speciﬁed upper bound for W problem for minimizing the total structural volume V (A) under a compliance constraint is formulated as Minimize

V (A) =

m ∑

Ai Li

(1.49a)

i=1

f (A) ≤ W U subject to W Ai ≥

ALi ,

(i = 1, . . . , m)

(1.49b) (1.49c)

30

Optimization of Finite Dimensional Structures

where Li is the length of member i, ALi is the lower bound for Ai , and the upper bound for Ai is not given for simplicity. The n × n stiﬀness matrix of a truss can be written as a linear function of A using constant n × n matrices Ki (i = 1, . . . , m) as K=

m ∑

Ai Ki

(1.50)

i=1

where

∂K (1.51) ∂Ai By using the last expression of (1.46) as well as (1.47) and (1.51), we obtain the sensitivity coeﬃcient (diﬀerential coeﬃcient) of compliance with respect to Ai as ( )⊤ f ∂W ∂W ∂U ∂W = + ∂Ai ∂Ai ∂U ∂Ai (1.52) ∂W = ∂Ai = −U⊤ Ki U Ki =

Although simpler derivations are possible for the sensitivity coeﬃcient, (1.52) is important for understanding the characteristics of the compliance. Let εi denote the strain of the ith member. Since U⊤ Ki U is twice the strain enf /∂Ai is equivalently ergy per unit cross-sectional area of the ith member, ∂ W written as f ∂W = −Eε2i Li (1.53) ∂Ai where E is the elastic modulus (see Sec. 2.2 for details of static sensitivity analysis). Alternatively, by diﬀerentiating (1.45) with respect to Ai and using (1.51), we obtain the sensitivity coeﬃcient of U with respect to Ai as ∂U = −K−1 Ki U ∂Ai

(1.54)

f (A) = U⊤ (A)K(A)U(A), the sensitivity coeﬃTherefore, using the form W cient of the compliance is given as f ∂W ∂U = U⊤ Ki U + 2U⊤ K ∂Ai ∂Ai = −U⊤ Ki U

(1.55)

which is the same as (1.52). Furthermore, diﬀerentiation of (1.45) with respect to Ai leads to ∂U =0 (1.56) Ki U + K ∂Ai

Various Formulations of Structural Optimization

31

By premultiplying U⊤ to (1.56), and using (1.45) and the ﬁrst expression of the compliance in (1.46), we obtain the same result as (1.52).

1.8.2

Optimality conditions

The Lagrange multiplier for the constraint (1.49b) is denoted by µ (≥ 0). If we consider the lower-bound constraint (1.49c) separately, the Lagrangian is formulated as ψ(A, µ) =

m ∑

f (A) − W U ) Ai Li + µ(W

(1.57)

i=1

Since P does not depend on A, the stiﬀness against P increases as Ai is f (A) generally decreases as A1 , . . . , Am are inincreased, and accordingly, W creased. Furthermore, V (A) is an increasing function of A1 , . . . , Am . Theref (A) ≤ W U is satisﬁed with equality at the optimal fore, the constraint W solution. Considering the side constraints for the variables, the necessary conditions for optimality are generally obtained using the derivatives of ψ(A, µ) as in (1.6); hence, from (1.53), we have { Eε2i = λ for Ai > ALi (1.58) Eε2i ≤ λ for Ai = ALi where λ = 1/µ > 0. It is seen from (1.58) that the absolute value of the strain, or the strain energy per unit volume, is the same for the members with Ai > ALi ; i.e., the optimal solution is fully stressed. This way, the optimality conditions are expressed in a simple form without sensitivity coeﬃcients of the displacements. Note that the conditions derived above using the Lagrangian are necessary conditions for optimality. In the following, we show, on the basis of the principle of minimum total potential energy, that the conditions (1.58) are also suﬃcient conditions for global optimality. Although suﬃcient conditions f (A), as demonstrated in Sec. 1.8.4, can be derived from the convexity of W we show below the conventional proof by Prager (1972). b = Suppose the optimality conditions (1.58) are satisﬁed by the solution A ⊤ ∗ ∗ ∗ ⊤ b b (A1 , . . . , Am ) . Let A = (A1 , . . . , Am ) denote a feasible solution that is b and has the same compliance value W U . The values corrediﬀerent from A b and A∗ are indicated by (c b εbi , sponding to A · ) and ( · )∗ , respectively, as U, ∗ ∗ U , εi , etc. Then we have U ∑ 1 ∑ ∗2 ∗ b ⊤P − 1 bi Li = W Eεi Ai Li = U Eb ε2i A 2 i=1 2 i=1 2 m

U∗⊤ P −

m

(1.59)

Since εb1 , . . . , εbm are the strains corresponding to an admissible displacement b that is not necessarily equal to the correct displacement vector U∗ vector U

32

Optimization of Finite Dimensional Structures

for A∗ , the following relation is obtained from the principle of minimum total potential energy: ∑ 1 ∑ ∗2 ∗ b ⊤P − 1 U Eb ε2i A∗i Li ≤ U∗⊤ P − Eεi Ai Li 2 i=1 2 i=1 m

m

(1.60)

By using (1.59), we obtain the following relation from (1.60): ∑ ∑ b ⊤P − 1 b ⊤P − 1 bi Li Eb ε2i A∗i Li ≤ U Eb ε2i A U 2 i=1 2 i=1 m

which is rewritten as

m ∑

m

bi Li )Eb (A∗i Li − A ε2i ≥ 0

(1.61)

(1.62)

i=1

From (1.62), we have m ∑

λ(A∗i

bi )Li − −A

i=1

m ∑

bi )(λ − Eb (A∗i − A ε2i )Li ≥ 0

(1.63)

i=1

bi = AL in the solution satLet I denote the set of indices of members with A i isfying the optimality conditions (1.58). Then, using (1.58), (1.63) is reduced to m ∑ ∑ bi )(λ − Eb bi )Li − (1.64) ε2i )Li ≥ 0 (A∗i − A λ(A∗i − A i=1

i∈I

bi = AL is satisﬁed for i ∈ I; therefore, from (1.58), we Furthermore, A∗i ≥ A i have ∑ bi )(λ − Eb (A∗i − A ε2i )Li ≥ 0 (1.65) i∈I

Finally, from (1.64), (1.65), and λ > 0, the following relation is derived: b = V (A∗ ) − V (A)

m ∑

bi )Li ≥ 0 (A∗i − A

(1.66)

i=1

Therefore, if (1.58) holds, no solution satisfying W (A∗ ) = W U has a smaller b which means that (1.58) is also a suﬃcient conobjective value than V (A), dition for global optimality. The necessary and suﬃcient conditions can also be obtained for problems with constraints on linear buckling load and eigenvalue of vibration in a similar manner using Rayleigh’s principle for the case where the associated eigenvalue problems have simple lowest eigenvalues (Prager 1974a).

Various Formulations of Structural Optimization L/2

33

L

(4) (1) 1 (2)

L/2

3

(3) 2 P1 , U1 P2 , U2

L

4 (5)

FIGURE 1.14:

A four-bar truss.

Example 1.9 Consider, as a small example, a four-bar truss as shown in Fig. 1.14. The lower-bound cross-sectional areas are given as ALi = A for all members. Let P1 = 2P and P2 = P , and deﬁne U1 and U2 as shown in Fig. 1.14. The total potential energy is given as Π=

A1 E 2 A2 E 2 A3 E 2 A4 E 2 U + U + U + U − 2P U1 − P U2 L 1 2L 1 L 2 2L 2

(1.67)

By diﬀerentiating Π with respect to U1 and U2 , respectively, we obtain the equilibrium equations as E E (2A1 + A3 )U1 = 2P, (2A2 + A4 )U2 = P L L

(1.68)

The sensitivity coeﬃcients of compliance are obtained as f f f f ∂W 2EU12 ∂ W EU12 ∂ W 2EU22 ∂ W EU22 =− , =− , =− , =− (1.69) ∂A1 L ∂A2 L ∂A3 L ∂A4 L Therefore, with the reciprocal λ of the Lagrange multiplier λ=

4EU12 4EU22 = 2 L L2

(1.70)

in (1.58), the following conditions are satisﬁed at the optimal solution: f f f f ∂W λL ∂W ∂W ∂W =− = , − ≤ λL, − ≤ λL ∂A1 ∂A3 2 ∂A2 ∂A4 A1 ≥ A, A2 = A, A3 ≥ A, A4 = A −

which leads to U1 = U2 =

L 2

√

λ f 3P L , W = E 2

√

λ E

(1.71a) (1.71b)

(1.72)

34

Optimization of Finite Dimensional Structures

Suppose W U and A are given with a coeﬃcient c as W U = cP L, A =

P 2Ec

(1.73)

Then we have U1 = U2 =

11P cL 5P 3P L , A1 = , A3 = , V = 3 4Ec 4Ec Ec

(1.74)

When the topology is optimized by letting A = 0, the displacements of the optimal solution are the same as those in (1.74), and the cross-sectional areas and the total structural volume are obtained as A1 =

3P 9P L 3P , A3 = , A2 = A4 = 0, V = Ec 2Ec 4Ec

(1.75)

Example 1.10 As a numerical example, an optimal solution is found for the 10-bar truss in Fig. 1.5 under a compliance constraint (Katoh, Ohsaki, and Tani 2002). The loads are given as (P1 , P2 ) = (0.0, 100.0 kN), and the upper bound for the compliance is 0.3 kNm. Other parameters are the same as those in Example 1.3 for optimization under stress constraints against a single loading condition. Optimal solutions are obtained using the optimization software package SNOPT Ver. 7.2 (Gill, Murray, and Saunders 2002). The optimal cross-sectional areas are similar to those in Fig. 1.6, where the values of A1 , . . . , A10 (mm2 ) are 1599.93, 800.06, 1131.27, 0.10, 0.10, 0.10, 799.94, 0.10, 1131.28, 0.10. The optimal objective value is 1.28005 × 107 mm3 . As is seen from Fig. 1.6, the members are located so as to eﬃciently transmit the load to the supports. Note that Ai = ALi is satisﬁed by the thin members, and the optimal solution is fully stressed, as is veriﬁed by the optimality conditions (1.58).

1.8.3

Reformulation of the optimization problem

As we have observed in Sect. 1.8.2, the optimal solution under a compliance constraint is equivalent to the fully-stressed design with σiL = −σiU for all members with appropriately chosen σiU , because the absolute values of stresses are the same for the members with Ai > ALi at the optimal solution. Furthermore, as we can intuitively recognize, minimization of compliance under volume constraint is equivalent, after an appropriate scaling of the objective function, to minimization of the volume under a compliance constraint, because the compliance is a decreasing function and the total structural volume is an increasing function of the cross-sectional areas. Therefore, utilizing these properties, many approaches have been presented for equivalent reformulation of the problem.

Various Formulations of Structural Optimization

35

Consider the following problem of minimizing the compliance under constraint on the total structural volume: f (A) = P⊤ U(A) W m ∑ subject to V (A) = Ai Li ≤ V U Minimize

(1.76a) (1.76b)

i=1

Ai ≥ ALi , (i = 1, . . . , m)

(1.76c)

where V U is the speciﬁed upper bound for the total structural volume. Let λ (≥ 0) be the Lagrange multiplier for the constraint (1.76b). With the use of (1.53), the Lagrangian without the side constraints (1.76c) and its sensitivity coeﬃcient with respect to Ai are written as ) (m ∑ U f Ai Li − V (1.77a) ψ(A, λ) = W (A) + λ i=1

∂ψ = −Eε2i Li + λLi , (i = 1, . . . , m) ∂Ai

(1.77b)

Therefore, from (1.77b) and the convexity of the compliance that is shown in Sec. 1.8.4, it is easily seen that the optimality conditions for Problem (1.76) are the same as (1.58), and Problem (1.76) is equivalent to Problem (1.49) if W U and V U are appropriately assigned (Achtziger 1997). For the four-bar truss in Example 1.9, 3P L VU = (1.78) Ec for Problem (1.76) leads to the same optimal solution as that of Problem (1.49). Consider the simple case where ALi = 0 for all members, which corresponds to the topology optimization problem that is extensively studied in Chap. 3. We can also use the approach of simultaneous analysis and design (SAND), as follows, with the member cross-sectional areas and the nodal displacements as variables (Bendsøe and Sigmund 2003) by explicitly incorporating the stiﬀness equation (1.45) that is reformulated using (1.50) into the constraints: P⊤ U m ∑ subject to Ai Ki U = P

Minimize

i=1 m ∑

Ai Li ≤ V U

(1.79a) (1.79b) (1.79c)

i=1

Ai ≥ 0, (i = 1, . . . , m) where the variables are A and U (see Sec. 1.13 for details of SAND).

(1.79d)

36

Optimization of Finite Dimensional Structures

A diagonal matrix S that consists of the extensional stiﬀness of each member is given as S = diag(EA1 /L1 , . . . , EAm /Lm ) (1.80) The vector of elongations of members is denoted by d = (d1 , . . . , dm )⊤ . Then the compliance is expressed as d⊤ Sd. Let D and N = (N1 , . . . , Nm )⊤ denote the equilibrium matrix and the vector of axial forces. The equilibrium equation DN = P and the constitutive relation N = Sd are given instead of the stiﬀness equation (1.79b). Hence, Problem (1.76) can be reformulated into the following form: Minimize d⊤ Sd subject to DN = P

(1.81a) (1.81b)

N = Sd m ∑ Ai Li ≤ V U

(1.81c) (1.81d)

i=1

Ai ≥ 0, (i = 1, . . . , m)

(1.81e)

where the variables are d, N, and A. Note that the compatibility conditions between d and U are ignored; however, they are satisﬁed at the optimal solution, as shown below. The Lagrangian for this problem is given as ψ = d⊤ Sd + µ⊤ (DN − P) + η ⊤ (N − Sd) (m ) m ∑ ∑ U +λ Ai Li − V − κi Ai i=1

(1.82)

i=1

where µ, η, λ (≥ 0), and κi (≥ 0) are the Lagrange multipliers. From the stationary conditions of the Lagrangian as well as the complementarity conditions κi Ai = 0, Ai ≥ 0, and κi ≥ 0, we have 2Sd − Sη = 0, D⊤ µ + η = 0 ∂S d⊤ d + λLi = 0 for Ai > 0 ∂Ai ∂S d + λLi ≥ 0 for Ai = 0 d⊤ ∂Ai

(1.83a) (1.83b) (1.83c) (1.83d)

We can see from these equations that the same optimality conditions as (1.58) are obtained, and there always exists a set of vectors U = (1/2)µ and d = (1/2)η satisfying the compatibility conditions (1.83b) at the optimal solution of Problem (1.81).

Various Formulations of Structural Optimization

37

Since the optimal solution for this problem is fully stressed, it can be obtained numerically in the same manner as the stress-ratio approach for fullystressed design. Let ei be deﬁned as ei =

1 ⊤ ∂S d d Li ∂Ai

(1.84)

which is twice the strain energy per unit volume of member i. The mean value of ei among the members with Ai > 0 is denoted by e. Then the cross-sectional areas are updated by ( ) (k) α ei VU (k+1) (k) = Ai Ai (1.85) V (A(k) ) e(k) where the superscript ( · )(k) denotes the value at the kth iteration, and α is a parameter that is slightly less than 1, e.g., 0.8 (Pedersen and Pedersen 2009). By expressing d with respect to N, we can formulate Problem (1.81) in terms of the axial force vector N (Achtziger 1997): Minimize

∑ Li N 2 i Ei Ai

(1.86a)

Ai ∈I /

subject to DN = P Ni = 0 for all i ∈ I m ∑ Ai Li ≤ V U

(1.86b) (1.86c) (1.86d)

i=1

Ai ≥ 0, (i = 1, . . . , m)

(1.86e)

where I is the set of indices of members satisfying Ai = 0. Let A denote the feasible region of A satisfying the constraints (1.79c) and (1.79d). Then Problem (1.79) can be rewritten using the last expression of (1.46) and the principle of minimum total potential energy as [ ) ] (m 1 ⊤ ∑ ⊤ (1.87) max min U A i Ki U − P U A∈A U 2 i=1 Because Problem (1.87) is linear with respect to A and convex with respect to U, global optimality is guaranteed for both problems of minimization with respect to U and maximization with respect to A. Therefore, the inner minimization problem and the outer maximization problem can be exchanged as (m ) ] [ 1 ⊤ ∑ ⊤ A i Ki U − P U min max U (1.88) U A∈A 2 i=1 In the conventional ground structure approach for topology optimization, ideally all the pairs of nodes should be connected by members that can exist;

38

Optimization of Finite Dimensional Structures

see Sec. 3.3. Hence, the number of components in A is far greater than that of U. Therefore, it is desirable to reformulate the problem to include only U as the variables. It is seen from the optimality condition (1.58) that (U⊤ Ki U)/Li , which is twice the strain energy per unit volume of member i, has the same value for the members with Ai > 0 at the optimal solution. We also have the relation m ∑ Ai Li = V U (1.89) i=1

because the volume constraint is satisﬁed with equality at the optimal solution. Therefore, the total strain energy of the optimal solution is written as V U (U⊤ Ki U)/(2Li ), where i represents an arbitrary member satisfying Ai > 0. Then, Problem (1.88) is reformulated with respect to U only as a variable vector (Ben-Tal and Bendsøe 1993): ( min max

U i=1,...,m

) VU ⊤ U Ki U − P⊤ U 2Li

(1.90)

Since the function in the parentheses in (1.90) is convex with respect to U, and the pointwise maximum of convex functions is also convex, as shown in Appendix A.1.1, Problem (1.90) is an unconstrained minimization problem of a convex function with respect to U. However, the objective function is nonsmooth, because we have to ﬁnd the maximum of m diﬀerent convex functions. The nonsmoothness of Problem (1.88) can be alleviated by using the standard reformulation with an arbitrary variable ν as Minimize

ν

(1.91a) U

subject to

V U⊤ Ki U − P⊤ U ≤ ν, (i = 1, . . . , m) 2Li

(1.91b)

which is alternatively written with an auxiliary variable τ (> 0) (Ben-Tal, Koˇcvara, and Zowe 1993): Minimize

τ − P⊤ U

(1.92a)

U

subject to

V U⊤ Ki U ≤ τ, (i = 1, . . . , m) 2Li

(1.92b)

Problem (1.92) is equivalently written as follows after scaling P and U by √ τ: Minimize

− P⊤ U

(1.93a)

U

subject to

V U⊤ Ki U ≤ 1, (i = 1, . . . , m) 2Li

(1.93b)

Various Formulations of Structural Optimization

39

Using the relation U⊤ Ki U = Eε2i Li in (1.53), Problem (1.93) can be reformulated in a form of fully-stressed design (Achtziger, Bendsøe, Ben-Tal, and Zowe 1992; Beckers and Fleury 1997; Achtziger 1997): Minimize

− P⊤ U √

subject to − 1 ≤

(1.94a) V U E b⊤ i U ≤ 1, (i = 1, . . . , m) 2 Li

(1.94b)

where b⊤ i deﬁnes the strain-displacement relation as εi =

b⊤ i U Li

(1.95)

Because Problem (1.94) is an LP problem with n variables, n inequality constraints in (1.94b) are satisﬁed with equality; i.e., the member stresses are obtained such that they take upper- or lower-bound values in n members. Then, from the optimality conditions, the cross-sectional areas of the remaining m − n members are 0, and those of m members are computed to satisfy the equilibrium equations. Hence, the optimal truss is statically determinate and fully stressed for the topology optimization problem with ALi = 0 for all members. Example 1.11 Consider again the simple four-bar truss, as shown in Fig. 1.9 in Example 1.14, with V U = 9P L/(4Ec). Let the term in the parentheses in (1.90) be denoted by fi , i.e., for this example, ( )( )( ) 9P L E 2 1 fi = d − 2P U1 − P U2 (1.96) 4Ec 2Li Li i From the relations L1 = L3 = L/2, L2 = L4 = L, d1 = −d2 = U1 , and d2 = −d4 = U2 , we can see that f1 > f2 and f3 > f4 are always satisﬁed. Therefore, minimization of the maximum of f1 and f3 leads to f1 = f3 , from which U1 = U2 is obtained. Incorporating U2 = U1 into (1.96), we have the stationary condition of f3 with respect to U1 as ( )( )( ) 9P L 2 2E (2U1 ) − 3P = 0 (1.97) 4Ec L L from which we obtain U1 = U2 = cL/3, which agrees with the results in (1.74). It is easily seen that the same solution can be found by solving Problems (1.92), (1.93), and (1.94), respectively.

1.8.4

Convexity of compliance

f (A) of a truss can be proved in several The convexity of the compliance W manners. The suﬃciency of the conditions (1.58) in Sec. 1.8.2 can be derived f (A). from the convexity of W

40

Optimization of Finite Dimensional Structures

The conventional approach for the proof of convexity utilizes the minimum principle of total potential energy: Proof 1.1 Let Π(A, U) denote the total potential energy that is conceived as a function of A and U as 1 Π(A, U) = U⊤ K(A)U − P⊤ U (1.98) 2 From the minimum principle of total potential energy, the correct U for the speciﬁed A is found by minimizing Π(A, U) with respect to U: U(A) = argmin Π(A, U)

(1.99)

U

f (A) = W (U(A), A) is which leads to KU = P. Therefore, the compliance W written as f (A) = −2 min Π(A, U) W (1.100) U

I

II

Let A and A denote two diﬀerent solutions. The displacement vectors for AI and AII are denoted by UI and UII , respectively. A linear combination of b = αAI + (1 − α)AII . AI and AII is given using a parameter 0 ≤ α ≤ 1 as A From the linearity of Π(A, U) with respect to A, which is easily seen from (1.50), we have b =W f (A) f (αAI + (1 − α)AII ) W [ ] = −2 min αΠ(AI , U) + (1 − α)Π(AII , U)

(1.101)

U

Since U cannot minimize Π(AI , U) and Π(AII , U) simultaneously in general, we have [ ] min αΠ(AI , U) + (1 − α)Π(AII , U) U [ ] (1.102) ≥ min αΠ(AI , UI ) + (1 − α)Π(AII , UII ) UI ,UII

Therefore, from (1.101) and (1.102), we obtain [ ] b ≤ −2 min αΠ(AI , UI ) + (1 − α)Π(AII , UII ) f (A) W UI ,UII

f (AI ) + (1 − α)W f (AII ) = αW

(1.103)

f (A). which shows the convexity of W Let In denote the n×n identity matrix. Diﬀerentiating the obvious relation KK−1 = In with respect to Ai , we obtain ∂K −1 ∂K−1 K +K =O ∂Ai ∂Ai

(1.104)

Various Formulations of Structural Optimization

41

From (1.50) and (1.104), the sensitivity coeﬃcient of K−1 with respect to Ai is expressed as ∂K−1 ∂K −1 = −K−1 K ∂Ai ∂Ai = −K−1 Ki K−1

(1.105)

Svanberg (1984, 1994) presented a proof based on the positive semideﬁf (A), assuming that W f (A) is twice continuously niteness of the Hessian of W diﬀerentiable with respect to A and K is nonsingular: Proof 1.2 f (A) = P⊤ U and K(A)U = P, the compliance can be rewritten From W without using U as f (A) = P⊤ K−1 (A)P W (1.106) where K−1 is a ﬂexibility matrix which is symmetric and positive deﬁnite. f (A) with respect to Ai and using (1.105), we obtain Diﬀerentiating W f ∂W ∂K−1 = P⊤ P ∂Ai ∂Ai = −P⊤ K−1 Ki K−1 P

(1.107)

Further diﬀerentiating with respect to Aj , and using U = K−1 P, the Hessian f (A) is obtained as of W f (A) ∂2W ∂K−1 ∂K−1 Ki K−1 P + P⊤ K−1 Ki P = −P⊤ ∂Ai ∂Aj ∂Aj ∂Aj = P⊤ K−1 Kj K−1 Ki K−1 P + P⊤ K−1 Ki K−1 Kj K−1 P

(1.108)

= U⊤ Kj K−1 Ki U + U⊤ Ki K−1 Kj U For an arbitrary vector y = (y1 , . . . , ym )⊤ , deﬁne the vector z as z=

m ∑

yi Ki U

(1.109)

i=1

Note that z is a kind of equivalent nodal load vector corresponding to U if yi is regarded as the cross-sectional area of the ith member. Then, we can show the following relation from (1.108) and symmetry of K−1 : ( ) m m f ∂2W 1 ∑∑ yi yj 2 i=1 j=1 ∂Ai ∂Aj =

m ∑ m ∑

(yj U⊤ Kj )K−1 (yi U⊤ Ki )

i=1 j=1

= z⊤ K−1 z

(1.110)

42

Optimization of Finite Dimensional Structures

Since K of a stable truss is positive deﬁnite, and accordingly, K−1 is positive f (A) is positive semideﬁnite. deﬁnite, (1.110) implies that the Hessian of W Note that the Hessian in (1.108) is not always positive deﬁnite, because there may exist a member that does not deform under displacement U; i.e., z may be a null vector for a nonzero vector y. This concludes the proof of convexity f (A). of W Svanberg (1994) also showed that the compliance is a concave function of the reciprocals of the cross-sectional areas, which ensures global convergence of the convex linearization method (CONLIN) (Nguyen, Strodiot, and Fleury 1987) applied to the compliance optimization problem. Finally, Stolpe and Svanberg (2001a) presented a proof based on the convexity of a function deﬁned as the pointwise maximum of a collection of convex functions: Proof 1.3 Let G(A, U) be deﬁned as (

1 G(A, U) = 2 P U − U⊤ K(A)U 2 Then, for every A,

⊤

)

f (A) = max G(A, U) W U

(1.111)

(1.112)

Suppose the extensional stiﬀness of member i is a concave function (including linear function) of Ai , which is valid for a truss. Then, for each ﬁxed U, G(A, U) is convex with respect to A. Because a function deﬁned as a pointf (A) is a convex wise maximum of a collection of convex functions is convex, W function (see Appendix A.1.1 for the deﬁnition of pointwise maximum). For a frame that has a solid section, the compliance is not generally a convex function of the cross-sectional areas; e.g., the bending stiﬀness of a section with constant height/width ratio is proportional to the square of the cross-sectional area with a positive coeﬃcient. Therefore, the stiﬀness matrix is a convex function of the cross-sectional parameters, and the requirement for Proof 1.3 is not satisﬁed.

1.8.5

Other topics on compliance optimization

As we have seen, the stiﬀest structure under speciﬁed static loads can be obtained through minimization of the compliance. By contrast, if forced displacements are applied, the work done by the reactions and the displacements should be maximized so as to ﬁnd the stiﬀest structure. For geometrically nonlinear problems, Buhl, Pedersen, and Sigmund (2000) deﬁned end compliance, which is the external work at the ﬁnal state. Contrary to the attempt to design the stiﬀest structure, optimization approaches have been developed to generate a compliant mechanism that has

Various Formulations of Structural Optimization

43

ﬂexibility against the applied loads to produce a speciﬁed large deformation. Nishiwaki, Min, Yoo, and Kikuchi (2001) used mutual energy for retaining stiﬀness at the output node against the input load, which is utilized by Saxena and Ananthasuresh (2000) for frame mechanisms. Ohsaki and Nishiwaki (2005) introduced stiﬀness constraints for the initial undeformed and the ﬁnal deformed state for generating a compliant bar-joint system that has multiple stable equilibrium states utilizing snapthrough behavior.

1.9

Frequency constraints

The fundamental natural frequency (or period) is an important performance measure of structures in the various ﬁelds of engineering, because it deﬁnes the stiﬀness against dynamic loads, e.g., earthquake loads, wind loads, in civil engineering. Furthermore, it is important that the higher-order frequencies are also appropriately assigned so as to prevent resonance to the possible dynamic loads. In this section, basic formulations are presented for optimization of trusses under frequency constraints. Detailed results of topology optimization under multiple eigenvalue constraints are shown in Sec. 3.9. Let Ms (A) and M0 denote the n × n mass matrices corresponding to the structural mass and nonstructural mass, respectively, with n being the number of degrees of freedom. Note that Ms (A) represents the mass of the structural members, which is a function of the vector A of the cross-sectional areas that are the design variables. In the following, the argument A is omitted for brevity. The eigenvalue problem of free vibration is formulated as KΦr = Ωr (M0 + Ms )Φr , (r = 1, . . . , n)

(1.113)

where Ωr and Φr are the rth eigenvalue and eigenmode, respectively, which are functions of A. The vector Φr is ortho-normalized by Φ⊤ r MΦs = δrs , (r = 1, . . . , n)

(1.114)

where δrs is the Kronecker delta. The objective here is to ﬁnd an optimal solution of minimizing the total structural volume under constraint such that the fundamental eigenvalue Ω1 is not less than the speciﬁed lower bound ΩL . Hence, the optimization problem is formulated as Minimize

V =

m ∑

Ai Li

i=1 L

subject to Ω1 ≥ Ω , (r = 1, . . . , n) Ai ≥

ALi ,

(i = 1, . . . , m)

(1.115a) (1.115b) (1.115c)

44

Optimization of Finite Dimensional Structures (1)

M0

1

FIGURE 1.15:

(2)

M0

2

A two-bar lumped mass structure.

where m is the number of members, ALi is the lower bound for Ai , and the upper bound for Ai is not given for simplicity. Deﬁne n × n matrices Ki and Mi as Ki =

∂K ∂Ms , Mi = ∂Ai ∂Ai

(1.116)

For a truss, we have K=

m ∑

A i Ki , M =

i=1

m ∑

Ai Mi + M0

(1.117)

i=1

If the fundamental eigenvalue is simple, the optimality conditions are derived using the Lagrange multiplier approach. The Lagrangian ignoring the side constraints (1.115c) is given as ψ(A, µ) =

m ∑

Ai Li + µ(ΩL − Ω1 )

(1.118)

i=1

where µ (≥ 0) is the Lagrange multiplier. Since the total structural volume is an increasing function of A, and Ω1 is generally an increasing function of A, the constraint (1.115b) is satisﬁed with equality assuming that suﬃciently small values are assigned for ALi . Therefore, incorporating the sensitivity coeﬃcients of Ω1 with respect to Ai presented in Sec. 2.3 to the KKT conditions (1.6), we obtain the following optimality conditions: { ⊤ Φ⊤ for Ai > ALi 1 Ki Φ1 − Ω1 Φ1 Mi Φ1 = Li /µ (1.119) ⊤ Φ⊤ for Ai = ALi 1 Ki Φ1 − Ω1 Φ1 Mi Φ1 ≤ Li /µ where µ > 0 and the normalization condition (1.114) has been used. Note ⊤ that Φ⊤ 1 Ki Φ1 and Ω1 Φ1 Mi Φ1 are conceived as twice the strain energy and kinetic energy, respectively, per unit area corresponding to Φ1 . Example 1.12 Consider a two-bar structure, as shown in Fig. 1.15. The two bars have the same elastic modulus E and length L, and the structural mass is ignored; i.e.,

Various Formulations of Structural Optimization

45

Ms = O. The stiﬀness matrix and mass matrix corresponding to nonstructural mass are given as ( ) ( ) E A1 + A2 −A2 M0 0 0 K= , M = (1.120) −A2 A2 0 M0 L The components of Φ1 are written as Φ1 = (Φ1,1 , Φ1,2 )⊤ . Then, from the optimality conditions, we have (Φ1,1 )2 = (Φ1,1 − Φ1,2 )2 = λ

(1.121)

with a positive constant λ. Then, using the normalization condition (1.114), we obtain ( ) 1 1 Φ1 = √ (1.122) 5M0 2 Hence, from (1.113) and (1.115b) satisﬁed with equality, the optimal crosssectional areas are obtained as A1 =

3M0 LΩL 2M0 LΩL , A2 = E E

(1.123)

From Rayleigh’s principle, the following inequality is satisﬁed by an arbitrary admissible n-vector Ψ (see Appendix A.1.2): Ψ⊤ KΨ ≥ Ω1 Ψ⊤ (Ms + M0 )Ψ

(1.124)

where the equality holds if Φ is the eigenmode corresponding to Ω1 . If Ω1 is a concave function of A, then (1.119) is the necessary and suﬃcient condition for global optimality. A proof of concavity is given below for the special case in which the structural mass is suﬃciently smaller than the nonstructural mass; i.e., Ms = O can be assumed. Proof 1.4 Let ΩI1 and ΩII 1 denote the lowest eigenvalues corresponding to the solutions I II A and A , respectively, for which other variables are also denoted by the superscripts I and II. The solutions are interpolated between AI and AII b = αAI + (1 − α)AII . The lowest eigenmode using a parameter 0 ≤ α ≤ 1 as A b b 1 . The following relations hold associated with the solution A is denoted by Φ from Rayleigh’s principle: b ⊤ KI Φ b1 Φ 1 b ⊤ M0 Φ b1 Φ 1

≥ ΩI1 ,

b ⊤ KII Φ b1 Φ 1 b ⊤ M0 Φ b1 Φ 1

≥ ΩII 1

(1.125)

46

Optimization of Finite Dimensional Structures

b 1 of Therefore, the following inequality is derived for the lowest eigenvalue Ω b A: I II b b⊤ b 1 = Φ1 [αK + (1 − α)K ]Φ1 Ω ⊤ b M0 Φ b1 Φ 1

=α ≥

b ⊤ KI Φ b1 Φ 1

+ (1 − α)

b ⊤ M0 Φ b1 Φ 1 αΩI1 + (1 − α)ΩII 1

b ⊤ KII Φ b1 Φ 1

(1.126)

b ⊤ M0 Φ b1 Φ 1

which proves that the lowest eigenvalue Ω1 is a concave function of A. Since Problem (1.115) is a standard nonlinear programming (NLP) problem, the optimal solution may be found by a gradient-based NLP algorithm in conjunction with the sensitivity analysis described in Sec. 2.3. However, it is well known that the lowest eigenvalue often becomes multiple as the result of optimization (Olhoﬀ and Rasmussen 1977; Olhoﬀ 1980; Haug, Choi, and Komkov 1986; Nakamura and Ohsaki 1988); see Sec. 3.9 for details. Therefore, the lower bounds should be given for all eigenvalues, and the optimization problem for a truss is formulated as Minimize

V =

m ∑

Ai Li

i=1 L

subject to Ωr ≥ Ω , (r = 1, . . . , n) Ai ≥

ALi ,

(i = 1, . . . , m)

(1.127a) (1.127b) (1.127c)

Although Problem (1.127) is also an NLP problem, the convergence of the algorithm is deteriorated due to the discontinuity of the sensitivity coeﬃcients; see Sec. 2.3.2 for details. The optimal solution with multiple eigenvalues can be found without any diﬃculty by converting Problem (1.127) to a semideﬁnite programming (SDP) problem (Wolkowicz, Saigal, and Vandenberghe 2000; Ohsaki, Fujisawa, Katoh, and Kanno 1999) (see Appendix A.2.4 for details of SDP). Using Rayleigh’s principle (1.124), the eigenvalue constraint is converted as Ψ⊤ [K − ΩL (Ms + M0 )]Ψ ≥ 0

(1.128)

for an arbitrary n-vector Ψ. Deﬁne an n × n matrix X as X = K − ΩL (Ms + M0 )

(1.129)

We can see from (1.128) and (1.129) that X should be positive semideﬁnite. Hence, the optimization problem (1.127) is reformulated as an SDP problem:

Various Formulations of Structural Optimization

L

47

L

(4) (3) M0

(2)

(1)

1 (5)

FIGURE 1.16:

H 3

2

H

A four-bar truss with a concentrated mass.

Minimize

V =

subject to X =

m ∑ i=1 m ∑

Ai Li

(1.130a)

(Ki − ΩL Mi )Ai − ΩL M0

(1.130b)

i=1 Ai ≥ ALi

X≽O

(i = 1, . . . , m)

(1.130c) (1.130d)

where X and A = (A1 , . . . , Am )⊤ are the variable matrix and vector, respectively, and X ≽ O means X is positive semideﬁnite (see Sec. 3.9 for details of the SDP formulation of the truss topology optimization problem under eigenvalue constraints). Example 1.13 As a simple example of the formulation of Problem (1.130), consider a fourbar truss, as shown in Fig. 1.16, for which a detailed investigation on multiple eigenvalues is presented in Sec. 3.9. The mass density of members, the lumped mass at node 3, and the elastic modulus are denoted by ρ, M0 , and E, respectively. The horizontal and vertical members have the same cross-sectional areas, respectively, denoted by A1 and A2 . Suppose the lumped mass matrix is used for representing the structural mass of the members. Then the matrix X in Problem (1.130) is deﬁned as 2E/L − ΩL ρL 0 −E/L A1 0 ρL 0 X= −E/L 0 E/L − ΩL ρL/2 L (1.131) −Ω ρH 0 0 0 0 0 2E/H − ΩL ρH 0 A2 − ΩL 0 0 0 + 0 0 0 M0 0 0 0

48

Optimization of Finite Dimensional Structures x 1 y

a

H

2 P H 3

W

FIGURE 1.17:

1.10

A three-bar truss.

Configuration optimization of trusses

So far we mainly considered the optimization problems of trusses with ﬁxed nodal locations, where the cross-sectional areas are the design variables. The topology of the truss can be optimized if zero lower bounds are assigned for the cross-sectional areas. By contrast, the nodal coordinates are modiﬁed in shape or geometry optimization. The eﬀectiveness of conﬁguration optimization, which combines shape and topology optimization, is illustrated below (Ohsaki 2003c); see Secs. 4.3, 4.4, and 6.6 for the numerical approaches. Example 1.14 As a simple example, consider a three-bar truss, as shown in Fig. 1.17, subjected to a vertical load P at node ‘a’. The distance of the node from the wall, on which three supports are located, has the ﬁxed value at W . The objective function to be minimized is the vertical displacement U of node ‘a’. The total structural volume is ﬁxed at V , and the three members have the same cross-sectional area A, for simplicity. Therefore, A can be computed from V as V A= (1.132) L1 + L2 + L3 where Li (i = 1, 2, 3) is the length of the ith member. Consider three cases in which H is equal to 2W , W , and W/2, respectively, and suppose the y-coordinate Y of node ‘a’ is chosen as the design variable for each case. Because L1 , L2 , and L3 depend on Y , the cross-sectional

Various Formulations of Structural Optimization

49

Displacement of node ‘a’

16 H = 2W H=W H = W/2 12

8

4 0

0.2 0.4 0.6 0.8

1

1.2 1.4 1.6 1.8

2

y-coordinate of node ‘a’

FIGURE 1.18: Relation between non-dimensional displacement U ∗ and non-dimensional nodal coordinate Y ∗ of the three-bar truss.

Displacement of node ‘a’

20

three members remove member 1 remove member 2

15 10 5 0 0

0.2 0.4 0.6 0.8

1

1.2 1.4 1.6 1.8

2

y-coordinate of node ‘a’

FIGURE 1.19: Relations between non-dimensional displacement U ∗ and non-dimensional nodal coordinate Y ∗ of the three-bar and two-bar trusses after removing a member.

area A computed from (1.132) is a function of Y , which is the only design variable. Fig. 1.18 shows the relations between Y and U for three cases with H = 2W , W , and W/2. Note that Y and U are converted to non-dimensional values Y ∗ = Y /H and U ∗ = EV U/(P W 2 ), respectively, with E being the elastic modulus. As is seen, U ∗ has two local minima with respect to Y ∗ for H = 2W , and U ∗ has the smallest value 5.4142 at Y ∗ = 1 for H = W . Hence, the nodal displacement under constraint on the total structural volume can be eﬀectively reduced through geometry optimization considering the location of a node as a design variable. Since this three-bar truss is statically indeterminate, the truss after removing one member is still stable. We consider the case with H = W , which has

50

Optimization of Finite Dimensional Structures x 1

H

y a 2 P

H

3

W

FIGURE 1.20: Global optimal shape and topology of the three-bar truss. the smallest objective value among the three cases. The relations between Y ∗ and U ∗ after removing members 1 and 2, respectively, are plotted in dashed and dotted lines in Fig. 1.19. As is seen, removal of member 2 leads to the smallest objective value U ∗ = 4.0000, and the corresponding global optimal topology is as shown in Fig. 1.20. This way, the objective value is drastically reduced by optimizing topology as well as the geometry.

1.11 1.11.1

Multiobjective structural optimization Basic concepts

In the design process of structures in civil engineering, several performance measures against dynamic loads (seismic loads, wind loads) and static loads (service loads, self-weight) should be considered. Furthermore, cost should be deﬁned in view of various aspects, including material, construction, manufacturing, and maintenance. Therefore, the optimization problem for practical application should be formulated as a multiobjective programming (MOP) problem (Cohon 1978) that has several performance measures as objective functions to be minimized or maximized. Multiobjective programming arises in the following situations in the design process: 1. The structural weight or volume is to be minimized while the struc-

Various Formulations of Structural Optimization

51

tural responses are reduced. In this case, the problem may be simply converted to a standard single-objective problem for minimizing the structural weight/volume under constraints on structural responses. 2. The initial cost as well as the maintenance cost should be minimized in the context of life-cycle design optimization; see Sec. 5.1.5. 3. The responses under various levels of seismic excitations are considered in the framework of performance-based design or performance-based engineering; see Sec. 5.5. 4. The mechanical performances as well as the aesthetic aspects and constructability should be considered in designing spatial frames and longspan structures; see Secs. 6.4 and 6.5. There are numerous papers published on MOP approaches to structural design, e.g., Koski and Silvennoisen (1987), Stadler (1988), Marler and Arora (2004). Ohsaki, Nakamura, and Isshiki (1998) presented a trade-oﬀ design approach for an arch-type truss considering seismic responses and the shape preferred by the designer. Ohsaki, Ogawa, and Tateishi (2003) developed a design method of shells considering roundness of the surfaces and stiﬀness against the static loads. Li, Zhou, Duan, and Chen (1999) presented an approach to multiobjective design of steel frames based on load and resistance factor design (LRFD) criteria. Ghasemi and Farshchin (2009) presented an ant colony optimization method for multiobjective optimization of steel frames considering structural weight and frequency of free vibration. One of the important aspects of MOP is that the competing performance measures can be formulated as either of the objective and constraint functions. The soft constraints that may preferably be satisﬁed can be naturally converted to an objective function (Ohsaki, Zhang, and Kimura 2005). Among the various methods of MOP, the constraint approach (Cohon 1978) and aspiration level approach (Nakayama 1995) for interactive optimization have no strict distinction between objective functions and constraints. The approach based on a fuzzy decision also treats objective and constraint functions equivalently (Rao 1987; Nakayama 1995; Dubois and Fortemps 1999; Massa, Lallemanda, and Tison 2009; Tonon 1999).

1.11.2

Problem formulation

For a simple case of minimizing the two objective functions F1 (A) and F2 (A), which are functions of the design variable vector A, the multiobjective optimization problem is formulated as Minimize F1 (A) and F2 (A) subject to A ∈ A

(1.133a) (1.133b)

52

Optimization of Finite Dimensional Structures

where A is the feasible region of A. If there exists no feasible solution that imb then A b is called the proves both F1 (A) and F2 (A) from a feasible solution A, Pareto optimal solution or simply the Pareto solution. The basic deﬁnitions and formulations of MOP are summarized in Appendix A.4. Example 1.15 As an illustrative example, consider a multiobjective structural optimization problem with discrete variables. Pareto optimal solutions are found for the ﬁve-bar truss shown in Fig. 1.2 in Sec. 1.3. Cross-sectional areas Ag1 and Ag2 of groups 1 and 2 consisting of members {1, 2, 5}, and {3, 4}, respectively, are selected from the list {100, 200, 300, 400, 500}; i.e., there are 5 × 5 = 25 possible combinations of design variables. Other parameters are the same as those in Example 1.1 in Sec. 1.3. Note that the units of length and force are mm and kN, which are omitted for brevity. The total structural volume and compliance are used as the performance measures to deﬁne the objective functions. The objective values of the 25 solutions are plotted in Fig. 1.21 in objective function space, where the ﬁlled circles indicate a Pareto optimal solution. The values of the objective functions and the design variables of the Pareto optimal solutions are also listed in Table 1.3. As is seen, the solutions with Ag1 = 500 or Ag2 = 100 are selected as Pareto solutions. This way, the Pareto solutions can be easily obtained by enumeration if the number of design variables is small, and the computational cost for function evaluation is also small. However, for more complex structures with many design variables, a nonlinear programming algorithm for a continuous problem or a heuristic approach for a discrete problem should be used, as presented in Secs. 5.4, 6.4, and 6.5. The solid line in Fig. 1.21 shows the set of Pareto solutions for continuous cross-sectional areas Ag1 and Ag2 with upper bound 500 and lower bound 100. In this example, the Pareto solutions of discrete variables are included in the set of continuous variables.

1.12

Heuristic approach

If the design variables are restricted to have only integer values, the design problem turns out to be a combinatorial optimization problem, for which integer programming approaches, such as the branch-and-bound method, can be eﬀectively applied. However, the computational cost grows as an exponential function of problem size. It should be noted, however, in the practical application, that obtaining the global optimal solution is not very important. Heuristic approaches have been developed to obtain approximate optimal so-

Various Formulations of Structural Optimization

53

Total volume (× 106 )

5 4 3 2 1 0 0

2

4

6

8

10

12

Compliance

FIGURE 1.21:

Pareto optimal solutions of the ﬁve-bar truss.

TABLE 1.3: Compliance and total structural volume of the Pareto optimal solutions of the ﬁve-bar truss with discrete cross-sectional areas. Compliance Ag2 Ag1 100.0 100.0 0.58363 200.0 100.0 0.35023 300.0 100.0 0.25678 400.0 100.0 0.20589 500.0 100.0 0.17379 500.0 200.0 0.14770 500.0 300.0 0.13394 500.0 400.0 0.12425 500.0 500.0 0.11673

Total volume (×106 ) 0.83006 1.2129 1.5957 1.9786 2.3614 2.8086 3.2559 3.7031 4.1503

lutions within reasonable computation times, although there is no theoretical proof of convergence to the optimal solution. The most popular heuristic approach may be the genetic algorithm (GA), which is classiﬁed as a multipoint method or population-based method that has many solutions at each iterative step. There are several diﬀerent terminologies for the classiﬁcation of the heuristic approaches. For example, a GA can be categorized as an evolutionary approach, a statistical search, a probabilistic approach, a soft computing method, or a method of artiﬁcial life. However, in this book, we use heuristic approach for all the approximate optimization methods without rigorous proof for reaching the global optimal solution. There are many papers on the application of GA to structural optimization

54

Optimization of Finite Dimensional Structures

(Ohsaki 1995; Hajela and Lee 1995). Schutte and Groenwold (2003) used a particle swarm optimization method for sizing optimization of trusses. Since the computational cost for function evaluation by structural analysis is not very small, a multipoint approach may not be appropriate for optimization of large-scale structures. Therefore, in this book, we mainly use heuristic approaches that are based on local search and retain only one solution at each iterative process. Such methods are called single-point-search heuristics. Balling (1991) applied simulated annealing (SA) to the optimization of building frames and showed that SA has better performance than a linearized branch-and-bound method. Lamberti (2008) proposed a multipoint SA and applied it to a truss optimization problem. Tagawa and Ohsaki (1999) used SA for conﬁguration optimization of trusses. Pan, Ohsaki, and Tagawa (2007) used SA for optimizing the ﬂange shape of a beam. The application of tabu search to structural optimization is found in Bennage and Dhingra (1995), Dhingra and Bennage (1995), Kargahi, Andersen, and Dessouky (2007), etc.; see Appendix A.3 for basic methodologies of heuristics. Example 1.16 As a small example, an optimal solution is found using the greedy method for the ﬁve-bar truss in Fig. 1.2 in Sec. 1.3. The problem formulation and parameters are the same as those in Sec. 1.3; i.e., the total structural volume V is minimized under constraints on stress and displacement, where the design variables Ag1 and Ag2 have discrete values {100, 200, 300, 400, 500}. In the following, the units for length and force are mm and kN, respectively. We start with the solution that has the smallest objective value and violates the constraints. Hence, we set the initial solution as (Ag1 , Ag2 ) = (100, 100) and carry out a local search by increasing either Ag1 and Ag2 , as illustrated with thick arrows in Fig. 1.22. The displacement and stress of the solution (Ag1 , Ag2 ) = (100, 100) are U3 = 2.362 and σ4 = −0.116, which do not satisfy the constraints |U3 | ≤ U 3 = 1.25 and |σ4 | ≤ σ 4 = 0.06. Deﬁne the parameter α as the maximum ratio of the constraint functions to their upper bounds; i.e., α is equal to the larger value between |U3 |/U 3 and |σ4 |/σ 4 . Therefore, the solutions are searched until α ≤ 1 is satisﬁed. At the initial solution, α = 1.93 and V = 8.30 × 105 . If Ag1 is increased by 100 to 200, then α = 1.44 and V = 1.21 × 106 . On the other hand, if Ag2 is increased to 200, then α = 1.50 and V = 1.27 × 106 . Let β denote the ratio of the decrease of α to the increase of V . Then it may be conceived that a larger value of β leads to more eﬃciency of the design modiﬁcation. The values of β for increasing Ag1 and Ag2 , respectively, in this case are 1.29 × 10−7 and 0.977 × 10−7 . Therefore, we increase Ag1 to obtain (Ag1 , Ag2 ) = (200, 100). However, α is still greater than 1, and Ag1 or Ag2 should be increased again. By comparing the values of β for updating the solution to (Ag1 , Ag2 ) = (300, 100) and (200, 200), we accept the latter solution, which

Various Formulations of Structural Optimization

55

500 400

U3 = const.

A2g

300 200

σ4 = const. 100 0 0

100

200

300

400

500

g 1

A FIGURE 1.22:

Local search of the ﬁve-bar truss.

satisﬁes the constraints. This way, the global optimal solution at point ‘b’ in Fig. 1.3 has been successfully found by the greedy method.

1.13

Simultaneous analysis and design

The structural optimization problem is usually formulated as Problem (1.3), where the nodal displacement vector U, which is also called state variable vector, is regarded as implicit function of the design variables A = (A1 , . . . , Am )⊤ . This type of problem formulation is sometimes called nested analysis and design, abbreviated NAND (Arora and Wang 2005), or design variable space approach (Kirsch and Rozvany 1994). In this case, the sensitivity of the constraint functions is evaluated using the sensitivity of U with respect to A by carrying out the standard procedure of design sensitivity analysis described in Chap. 2. However, for a complex structure, the sensitivity analysis is very tedious and demands much eﬀort in programming and computation. As an alternative approach, the nodal displacements are also considered as independent variables, while the stiﬀness equations are explicitly assigned as equality constraints. This combined approach of analysis and design is called simultaneous analysis and design (SAND) (Haftka 1985; Sankaranarayanan, Haftka, and Kapania 1994; Wu and Arora 1987), and the problem of minimizing the objective function F (A) under inequality constraints Hj (U, A) ≤ 0

56

Optimization of Finite Dimensional Structures

(j = 1, . . . , nI ) is formulated as Minimize

F (A)

(1.134a)

subject to Hj (U, A) ≤ 0, (j = 1, . . . , n ) K(A)U = P

(1.134b) (1.134c)

ALi ≤ Ai ≤ AU i , (i = 1, . . . , m)

(1.134d)

I

L where AU i and Ai are the upper and lower bounds for Ai , K is the stiﬀness matrix, and P is the speciﬁed vector of external loads. Using the SAND approach, optimal solutions can be found without carrying out structural analysis in a standard manner, and no procedure of design sensitivity analysis is needed. Although the number of variables increases, the nonlinearity of the governing equations is reduced, and the convergence properties may be improved. Taleb-Agha and Nelson (1975) optimized a truss with stress and displacement constraints under multiple loading conditions using SAND and convex linearization approaches. Wu and Arora (1987) presented a SAND formulation for a geometrically nonlinear problem. Shin, Haftka, and Plaut (1987) developed a method for eigenvalue optimization including bimodal cases. Orozco and Ghattas (1997) developed a reduced SAND approach utilizing sequential (recursive) quadratic programming for geometrically nonlinear structures. The attempt to solve the analysis problem by optimization was initiated in the 1960s (Schmit and Fox 1965). Note that we can carry out plastic limit analysis utilizing a linear programming approach, as demonstrated in Sec. 1.4. However, for a linear analysis problem, the standard approach of solving equilibrium equations is more eﬃcient than the optimization approach for minimizing the total potential energy. By contrast, for a geometrically nonlinear analysis and design problem, the computational cost can be reduced by utilizing the SAND formulation (Haftka and Kamat 1989). The internal forces, such as axial forces and stresses, can also be treated as independent variables (Wang and Arora 2006a, 2006b). For example, the equilibrium equations and compatibility conditions of a truss are incorporated as equality constraints to formulate the problem in the following form:

Minimize

F (A)

(1.135a)

subject to Hj (U, A) ≤ 0, (j = 1, . . . , nI ) DN = P ⊤

(1.135b) (1.135c)

S(A)D U = N

(1.135d)

ALi ≤ Ai ≤ AU i , (i = 1, . . . , m)

(1.135e)

where N is the vector of member forces, D is the equilibrium matrix, and S is the diagonal matrix of which the ith diagonal term is EAi /Li with E and Li being the elastic modulus and the length of the ith member, respectively.

Various Formulations of Structural Optimization

57

Recently, the formulation of SAND has been extensively used for reformulation of topology optimization problems of plates and trusses (Bendsøe and Sigmund 2003), as we have seen in Sec. 1.8.3. On the basis of the SAND formulation, optimal topologies of large-scale trusses can be eﬃciently found utilizing an interior-point method (Jarre, Koˇcvara, and Zowe 1998). Because the number of variables is very large and large-scale linear equations should be solved at each step of optimization, the use of an iterative solver with multigrid smoothing may be very eﬀective (Maar and Schulz 2000). The optimization problem with equilibrium constraints (1.134c) is generally categorized as a mathematical program with equilibrium constraints (MPEC) (Luo, Pang, and Ralph 1996).

Chapter 2 Design Sensitivity Analysis

As we have seen in various formulations of structural optimization in Chap. 1, the sensitivity coeﬃcients of responses with respect to the design variables are needed for obtaining optimal designs using a gradient-based nonlinear programming approach. In this chapter, we ﬁrst present an overview of sensitivity analysis in Sec. 2.1. The static and eigenvalue sensitivity analyses are presented in Secs. 2.2 and 2.3, respectively. Methods for linear buckling load, transient response, and geometrically nonlinear response are brieﬂy presented in Secs. 2.4, 2.5, and 2.6, respectively. Finally, shape sensitivity analysis of trusses is presented in Sec. 2.7.

2.1

Overview of design sensitivity analysis

The term sensitivity analysis is mathematically deﬁned as the procedure for computing the rate of variation of the solution with respect to a parameter deﬁning the problem. Consider a problem of ﬁnding a solution to a set of n nonlinear equations deﬁned as follows with a scalar parameter p: G(X, p) = 0

(2.1)

where X = (X1 , . . . , Xn )⊤ is the variable vector, and G(X, p) is supposed to be continuously diﬀerentiable with respect to X and p. Since the solution is deﬁned by (2.1) for each speciﬁed value of p, it is regarded as a function of e e 0 ) for p that is denoted with tilde as X(p). Suppose we have a solution X(p 0 0 p = p . Then the solution for p = p + ∆p can be linearly approximated as ( ) e 0 + ∆p) − X(p e 0 ) + ∂G ∆p = 0 B X(p (2.2) ∂p where the n × n matrix B is the Jacobian of G, for which the (i, j)-component Bij is deﬁned as ∂Gi (2.3) Bij = ∂Xj According to the implicit function theorem, (2.2) has a solution if B is nonsingular. Fig. 2.1 illustrates the process of linear estimation of the solution

59

60

Optimization of Finite Dimensional Structures

Xi

linear estimation using a sensitivity coefficient

p0 FIGURE 2.1:

p0 + ∆p

p

Linear estimation of solution using a sensitivity coeﬃcient.

with respect to the parameter variation by ∆p. This approach is generally used for path-tracing analysis of geometrically nonlinear responses, where the load factor, a speciﬁc displacement component, or arc-length of the path is taken as the path parameter. For the structural design problem, the displacements and stresses under static and/or dynamic loads depend on the design variables, including crosssectional areas and nodal coordinates of the structure. The ratio of variation of the response to that of the design variable is called the design sensitivity coeﬃcient, and the process of computing the design sensitivity coeﬃcient is called design sensitivity analysis (Haug, Choi, and Komkov 1986; Adelman and Haftka 1986; Haftka and Adelman 1989; Choi and Kim 2004; van Keulen, Haftka, and Kim 2005). For the shape design, the variables are nodal coordinates, and the corresponding sensitivity coeﬃcients and the process of computing them are called shape sensitivity coeﬃcients and shape sensitivity analysis, respectively. Various formulations of design sensitivity analysis have been developed mainly for continua utilizing variational principles, for which the sensitivity coeﬃcients are to be computed for a functional, e.g., compliance, that can be formulated in an integral of a bilinear form of the responses. However, for ﬁnite-dimensional structures, including frames and trusses, the variational formulations are not needed, and the sensitivity coeﬃcients can be found simply by diﬀerentiating the governing equations in matrix-vector form. Furthermore, even for the shape sensitivity analysis of continua such as plates and shells, it has been shown that diﬀerentiation of the matrix-vector form after discretization to ﬁnite elements leads to the same results obtained by diﬀerentiation in variational form followed by discretization to ﬁnite elements (Choi and Twu 1988). The process of structural design may be classiﬁed into the steps of determining support locations, support conditions, nodal locations, member locations, and cross-sectional properties of members. The steps up to the determina-

Design Sensitivity Analysis 2P

P 4

(2) Y

FIGURE 2.2:

3

1 X

2 (1)

61

(4) 5

(3)

7 6 (5)

A seven-bar plane truss.

tion of member locations are classiﬁed as shape design, and the last step is called stiﬀness design. As an introduction to this chapter, the role of design sensitivity analysis in structural analysis and design is discussed using the simple seven-bar truss, as shown in Fig. 2.2. Consider a static analysis problem for computing the responses against the nodal loads. Let ui and vi denote the displacements in the X- and Y -directions, respectively, of node i. Then the deformation of the truss is deﬁned by the nodal displacement vector U = (u2 , v2 , u3 , v3 , u4 , v4 )⊤ , which is also called the state variable vector. Because U varies with respect to the nodal coordinates and the member cross-sectional areas, it is conceived as a function of these variables. Suppose the Y -directional displacement v3 of node 3 exceeds the allowable limit deﬁned by the design code. The displacement may be reduced by increasing the cross-sectional areas of the members. The most eﬀective design modiﬁcation may be obtained accurately by carrying out structural analysis for all possible modiﬁcations. However, this process demands much computational cost. Therefore, it is practically desirable to estimate the displacement linearly utilizing the sensitivity coeﬃcients. As we have seen in Chap. 1.1, sensitivity analysis plays a key role in structural optimization. The sensitivity coeﬃcients are also eﬀectively used to reduce the computational costs for a hom*ology design (Yoshikawa, Elishakoﬀ, and Nakagiri 1998) or inverse design for controlling the displacements to a speciﬁed values or ratios; e.g., equal vertical displacements can be attained at nodes 2 and 4 of the truss in Fig. 2.2 by adjusting the cross-sectional areas of members. It is well known for frame structures that the use of reciprocals of variables leads to more accurate linear estimation of responses than the case in which the cross-sectional areas are directly used as design variables (Noor and Lowder 1975; Storaasli and Sobieszczanski-Sobieski 1983; Haftka, G¨ urdal, and Kamat 1990). Higher-order sensitivity coeﬃcients may be incorporated for more accurate estimation (Fleury 1989b; Fuchs 1993). Polynomial ﬁtting algorithms may also be used for obtaining approximate responses (Haftka, Nachlas, Watson, Rizzo, and Desai 1987).

62

Optimization of Finite Dimensional Structures 5 Y

Node 2 2

X

4

ti 6

Node 1

1

Z 3

FIGURE 2.3: Deﬁnition of coordinates, displacements, node numbers, and direction of a three-dimensional truss member.

2.2 2.2.1

Static responses Direct diﬀerentiation method

In order to present the basic idea of design sensitivity analysis, the most standard approach of the direct diﬀerentiation method for static responses of a three-dimensional truss is presented in this section. Let Ai and Li denote the cross-sectional area and length of the ith member, respectively. The global coordinates (X, Y, Z), member direction, local node numbers, and local displacement numbers are deﬁned in Fig. 2.3. Let ti denote the unit vector directed from node 1 to 2 of the ith members; i.e., the three components of ti correspond to the directional cosines with respect to the X-, Y -, and Z-axes, respectively. A vector d is deﬁned as ⊤ ⊤ di = (−t⊤ i , ti )

(2.4)

The nodal displacement vector and nodal force vector of member i with respect to the global coordinates are denoted by ui and fi , respectively, which have six components. Then ui and fi are related by the 6 × 6 member stiﬀness matrix ki with respect to the global coordinates as fi = ki u i

(2.5)

where ki =

Ai E di d⊤ i Li

(2.6)

with E being the elastic modulus. Then ki of all members are assembled to construct the n × n global stiﬀness matrix K, where n is the total number of degrees of freedom. For design sensitivity analysis, the cross-sectional areas A = (A1 , . . . , Am )⊤ are considered as design variables, where m is the number of members. The

Design Sensitivity Analysis

63

stiﬀness matrix is a function of A, which is written as K(A). If the selfweight is considered, the nodal load vector P = (P1 , . . . , Pn )⊤ is a function of A, which is written as P(A). The displacement vector U(A) = (U1 , . . . , Un )⊤ is obtained by solving the stiﬀness (equilibrium) equation K(A)U(A) = P(A)

(2.7)

Diﬀerentiation of (2.7) with respect to Ai leads to ∂K ∂U ∂P U+K = ∂Ai ∂Ai ∂Ai

(2.8)

In the following, the argument A is omitted for brevity. By diﬀerentiating (2.6) with respect to Ai , we obtain E ∂ki = di d⊤ i ∂Ai Li

(2.9)

which is constant and corresponds to the value of ki for unit cross-sectional area. Hence, the sensitivity coeﬃcients of K(A) with respect to Ai are easily obtained through an arithmetic operation, because K is an assemblage of ki . The sensitivity of P(A) is also easily computed, if the dependence of P(A) on A is deﬁned through the self-weight that is an explicit linear function of A. Deﬁne R as ∂P ∂K R= − U (2.10) ∂Ai ∂Ai Then, using (2.8) and (2.10), the sensitivity coeﬃcient of U with respect to Ai is obtained from ∂U K =R (2.11) ∂Ai As is seen from (2.7) and (2.11), the sensitivity vector ∂U/∂Ai is computed as the displacement vector against the nodal load vector R. The stiﬀness matrix K is decomposed to triangular and diagonal matrices by a Cholesky decomposition in the process of computing U from (2.7), if a direct solver of linear equations is used. Therefore, the computational cost of solving (2.11) is very small except for the case in which a recursive method, e.g., the conjugate gradient method, is used for solving the linear equations. Example 2.1 Consider a two-bar truss, as shown in Fig. 2.4, subjected to horizontal and vertical loads P1 and P2 , respectively. The nodal displacements are denoted by U1 and U2 . The two bars have the cross-sectional areas A1 and A2 . For the case where P1 = 0 and P2 = P , (2.7) is written as ( )( ) ( ) √ E U1 0 A1 + 2 2A2 A1 √ = (2.12) U P A A 2 1 1 2 2L

64

Optimization of Finite Dimensional Structures

1

L

A1 2

A2

P1 , U1

P2 , U 2

L

FIGURE 2.4:

A two-bar truss.

from which we obtain

√ (A1 + 2 2A2 )P L PL , U2 = U1 = − A2 E A1 A2 E

(2.13)

Let A2 be a design variable, and suppose that P is independent of A2 . Eq. (2.8) is then written as ∂U1 ( ) ( )( ) ( ) √ E 1 0 E U1 0 ∂A2 A1 + 2 2A2 A1 √ + = (2.14) U2 0 A A ∂U L 0 0 1 1 2 2 2L ∂A2 By incorporating (2.13) into (2.14), we obtain ∂U2 PL PL ∂U1 =− 2 , = 2 ∂A2 A2 E ∂A2 A2 E

(2.15)

which agrees with the results of explicit diﬀerentiation of U1 and U2 in (2.13) with respect to A2 . The sensitivity coeﬃcients can also be computed using the ﬁnite-diﬀerence approach. Let ∆Ai denote the small positive variation of Ai , and denote by ∆Ai the m-vector of which the ith component is ∆Ai and other components are 0. An approximate sensitivity coeﬃcient of U(A) is obtained by the central ﬁnite-diﬀerence approach as U(A + ∆Ai ) − U(A − ∆Ai ) ∂U ≃ ∂Ai 2∆Ai

(2.16)

An approximate sensitivity coeﬃcient can also be found by the forward ﬁnitediﬀerence U(A + ∆Ai ) − U(A) ∂U ≃ (2.17) ∂Ai ∆Ai

Design Sensitivity Analysis

65

or the backward ﬁnite-diﬀerence U(A) − U(A − ∆Ai ) ∂U ≃ ∂Ai ∆Ai

(2.18)

Although the ﬁnite-diﬀerence approaches are very simple, the number of response analyses (solution processes of (2.7)) for sensitivity evaluation with respect to m design variables is m for the forward or backward ﬁnite-diﬀerence approach, and 2m for the central ﬁnite-diﬀerence approach. Therefore, the computational cost for sensitivity analysis is drastically reduced if direct differentiation of the governing equations is utilized. Furthermore, there are some inherent sources of errors in ﬁnite-diﬀerence approaches. Suppose a function is approximated by a polynomial of a variable. Then, the ﬁnite diﬀerence approximation assumes that the increment of the function consists of the ﬁrst-order term only, and the second- and higher-order terms are ignored. This type of error is called truncation error and can be reduced using a small increment of a variable. However, a too small value of increment leads to underﬂow of the diﬀerences, and accuracy deteriorate due to rounding error (see for details, e.g., Forsythe and Wasow 1960; Richtmyer and Morton 1967). Therefore, it is not easy to decide on an appropriate value of the increment of the variable. For plates and shells discretized to ﬁnite elements, it is not straightforward to diﬀerentiate the element stiﬀness matrix, especially with respect to the nodal coordinates. The computational cost can be reduced, while the diﬀerentiation of the stiﬀness matrix is avoided, if a ﬁnite-diﬀerence approach is used only for the element-level diﬀerentiation, and (2.11) for the direct diﬀerentiation method is solved in the system level to obtain the sensitivity coeﬃcients of the displacements. This approach is called the semi-analytical approach (de Boer and van Keulen 2000) and is very eﬀective for shape sensitivity analysis of complex structures with a large number of degrees of freedom. Suppose that the sensitivity coeﬃcients are to be found for the local responses of members. Let εj and Nj denote the strain and axial force of the jth member that are deﬁned as 1 ⊤ d uj Lj i Aj E ⊤ Nj = d uj Lj i εj =

(2.19a) (2.19b)

Since uj consists of the components of U(A), the axial force is deﬁned by the displacements U(A) and the design variable Aj . Hence, a local response is generally deﬁned as D(U(A), A), which is alternatively written as a function of A only: e D(A) = D(U(A), A) (2.20)

66

Optimization of Finite Dimensional Structures

By diﬀerentiating (2.20) with respect to Ai , we obtain e ∂D ∂D ∑ ∂D ∂Uj = + , (i = 1, · · · , m) ∂Ai ∂Ai j=1 ∂Uj ∂Ai n

(2.21)

Because D is an explicit function of U and A, as is the case of Nj in (2.19b), the partial diﬀerential coeﬃcients ∂D/∂Ai and ∂D/∂Uj can be easily come can be obtained by incorporating puted, and the sensitivity coeﬃcient of D those of displacements into (2.21). For εj and Nj , the following equations are derived by diﬀerentiating (2.19a) and (2.19b) with respect to Ai : 1 ⊤ ∂uj ∂εj = d ∂Ai Lj i ∂Ai Aj E ⊤ ∂uj d ∂Nj Lj i ∂Ai = Ai E ⊤ ∂ui ∂Ai E d⊤ d i ui + Li Li i ∂Ai

(2.22a) for i ̸= j (2.22b) for i = j

Note in (2.22b) that there exist two terms for i = j due to explicit and implicit dependence on Ai , respectively, and only implicit dependence exists for i ̸= j. Since the computational cost in (2.21) is negligibly small, the total computational cost is governed by the cost for solving (2.11), which is proportional to the number of design variables. Therefore, the direct diﬀerentiation method is very eﬀective when the sensitivity coeﬃcients of many response quantities are needed with respect to a relatively small number of design variables. However, it is not eﬃcient if the number of design variables is far larger than that of the responses, for which the sensitivity coeﬃcients are to be computed. In such a case, the adjoint variable method presented in the following section is very eﬀective.

2.2.2

Adjoint variable method

In the adjoint variable method, the sensitivity coeﬃcients of the responses are found without computing those of all the nodal displacements. Our pure pose here is to compute the sensitivity coeﬃcients of the response D(A) = D(U(A), A), which represents the strain, stress, axial force of a member, etc. For this purpose, the adjoint variable vector U is ﬁrst introduced as KU = where ∂D = ∂U By premultiplying U

⊤

(

∂D ∂U

∂D ∂D ,··· , ∂U1 ∂Un

(2.23) )⊤ (2.24)

on both sides of (2.11), we obtain ⊤

U K

∂U ⊤ =U R ∂Ai

(2.25)

Design Sensitivity Analysis

67

The following equation is also derived by premultiplying (∂U/∂Ai )⊤ to (2.23): (

∂U ∂Ai

)⊤

( KU =

∂U ∂Ai

)⊤

∂D ∂U

(2.26)

Furthermore, from (2.25), (2.26), and symmetry of K, (

∂U ∂Ai

)⊤

∂D ⊤ =U R ∂U

(2.27)

is derived, and ﬁnally from (2.21), (2.20), and (2.27), we obtain e ∂D ∂D ⊤ = +U R ∂Ai ∂Ai

(2.28)

Therefore, after solving (2.23) for U, the sensitivity coeﬃcients of the response are obtained from (2.28) by a simple arithmetic operation. Note that ∂D/∂Ai corresponds to diﬀerentiation of the explicit term of Ai in D. This term vanishes for the responses such as displacements and strains, which are expressed directly using U only. Consider a simple case of D = Uk ; i.e., the sensitivity coeﬃcients of the kth displacement component are to be found. In this case, ∂D/∂U is a vector consisting of 0 in all components except 1 in the kth component, as is seen from (2.24). Therefore, from (2.23), U is the displacement vector against the k unit load in the kth component, which is denoted by U . Because A does not exist explicitly in the deﬁnition D = Uk , the following relation is obtained e = D = Uk : from (2.10), (2.28), and D ∂Uk k⊤ =U ∂Ai

(

) ∂P ∂K − U ∂Ai ∂Ai

(2.29)

Therefore, the sensitivity coeﬃcient of the displacement component can be obtained easily from (2.29) through a simple arithmetic operation. Since the computational cost is dominated by the computation of the adjoint variable vector in (2.23), the total computational cost is proportional to the number of response quantities for which the sensitivity coeﬃcients are to be computed. Therefore, this approach is useful if the number of design variables is far larger than that of the response quantities. We present below an explicit formulation of the adjoint variable method to compute the sensitivity coeﬃcients of the kth displacement component of a plane truss. The relation between the local displacement vector ui and the strain εi of the ith member is deﬁned in (2.19a). From (2.6), the sensitivity coeﬃcient of the member stiﬀness matrix ki is obtained as (2.9). Assume, for

68

Optimization of Finite Dimensional Structures

simplicity, that P is independent of Ai . Then (2.29) is reduced to ∂Uk k⊤ ∂K = −U U ∂Ai ∂Ai ∂ki = −uk⊤ ui i ∂Ai E = − uk⊤ di d⊤ i ui Li i

(2.30)

where uki is the local displacement vector of the ith member corresponding to k k U . The strain and axial force of member i corresponding to Ui are denoted k by εki and N i , respectively. Then (2.30) is rewritten as ∂Uk = −ELi εki εi ∂Ai Li k = − 2 N i Ni Ai E

(2.31)

Therefore, the sensitivity coeﬃcients of the kth displacement component can k be found easily from the strain εki or the axial force N i against the unit virtual load at the kth displacement component. Example 2.2 The sensitivity coeﬃcient of U2 is computed for the two-bar truss in Fig. 2.4. Eq. (2.23) is written as ) ( ) ( √ E 2 0 A1 + 2 2A2 A1 √ U = (2.32) 1 A1 A1 2 2L from which we obtain

L A2 E 2 U = (A1 + 2√2A2 )L A1 A2 E

Furthermore, from (2.10) and (2.13),

−

EU1 L R= 0 −

(2.33)

(2.34)

is derived. Then, from (2.13), (2.28), and (2.34), we obtain U1 PL ∂U2 =− =− 2 ∂A2 A2 A2 E which is the same as (2.15).

(2.35)

Design Sensitivity Analysis

2.3 2.3.1

69

Eigenvalues of free vibration Simple eigenvalue

We next present the methods of sensitivity analysis of eigenvalues and eigenmodes of free vibration. Let M(A) denote the n × n mass matrix of the structure, which is a function of the design variable vector A. The eigenvalue problem of free vibration is formulated as KΦr = Ωr MΦr , (r = 1, . . . , n)

(2.36)

where Ωr and Φr are the rth eigenvalue and eigenmode, respectively, which are functions of A. The eigenmode Φr is ortho-normalized with respect to M as Φ⊤ (2.37) r MΦs = δrs , (r, s = 1, . . . , n) where δrs is the Kronecker delta. Consider ﬁrst the case where Ωr is simple; i.e., Ωr−1 < Ωr < Ωr+1 is satisﬁed. Diﬀerentiation of both sides of (2.36) with respect to Ai reads ∂K ∂Φr ∂Ωr ∂M ∂Φr Φr + K = MΦr + Ωr Φr + Ω r M ∂Ai ∂Ai ∂Ai ∂Ai ∂Ai

(2.38)

Premultiplying Φ⊤ r on both sides of (2.38), using (2.36), (2.37), and the symmetry of M and K, we obtain ( ) ∂K ∂M ∂Ωr = Φ⊤ − Ω Φr (2.39) r r ∂Ai ∂Ai ∂Ai Note that K and M are explicit functions of Ai . Therefore, when the eigenvalue Ωr and the eigenmode Φr are known, the sensitivity coeﬃcients of eigenvalues are computed from (2.39) through a simple arithmetic operation without resort to the sensitivity coeﬃcients of the eigenmodes. In civil and architectural engineering, the natural frequency and natural period are usually used rather than the eigenvalue. After √obtaining the sensitivity coeﬃcients of Ω , those of the rth frequency f = Ωr /(2π) and period r r √ Tr = 2π/ Ωr are computed as 1 ∂Ωr ∂Tr π ∂Ωr ∂fr = , = −√ 2 ∂Ai 8π fr ∂Ai ∂Ai Ω3r ∂Ai

(2.40)

Next, we carry out sensitivity analysis of eigenmodes (Yu, Liu, and Wang 1996). Diﬀerentiation of (2.37) with respect to Ai leads to Φ⊤ r

∂M ∂Φr Φr + 2Φ⊤ =0 r M ∂Ai ∂Ai

(2.41)

70

Optimization of Finite Dimensional Structures

From (2.38) and (2.41), ) ( ∂Φr ∂M ∂K ( ) Ωr ∂Ai − ∂Ai Φr K − Ωr M −MΦr ∂Ai = ⊤ ∂Ωr −Φr M 0 1 ⊤ ∂M Φr Φr ∂Ai 2 ∂Ai

(2.42)

is derived. Therefore, the sensitivity coeﬃcients of an eigenmode can be found by solving the set of n + 1 simultaneous linear equations (2.42). However, it is not straightforward to see that (2.42) has a unique solution. For the case of a simple eigenvalue, the rank of K − Ωr M is n − 1 with the kernel (null-space) spanned by Φr . Let C denote the matrix on the left-hand side of (2.42). Because the vector MΦr generally has the component of Φr , the n × (n + 1) matrix B, deﬁned as follows, has full row rank that is equal to n: ( ) B = K − Ωr M MΦr (2.43) Furthermore, the vector (−Φ⊤ r M 0) in the (n + 1)st row of C cannot be expressed as a linear combination of the rows of the matrix B, because −Φ⊤ r M has a component of the kernel of K−Ωr M. Hence, the matrix C has full rank n + 1, and (2.42) has a unique solution. Therefore, the sensitivity coeﬃcients of eigenvalues and eigenmodes can be simultaneously computed from (2.42). Suppose the sensitivity coeﬃcient of an eigenvalue has already been computed using (2.39), and deﬁne a vector b as b=

∂Ωr ∂M ∂K MΦr + Ωr Φr − Φr ∂Ai ∂Ai ∂Ai

(2.44)

If b is orthogonal to Φr , i.e., if b does not belong to the kernel of K − Ωr M, then the ﬁrst equation in (2.42) has a solution. Furthermore, it can be easily conﬁrmed that the orthogonality of b to Φr leads to the relation (2.39); hence, b is actually orthogonal to Φr , and it is also possible to obtain the sensitivity coeﬃcients of the eigenmodes after computing those of the eigenvalues. In this approach, however, the coeﬃcient matrix on the left-hand side of (2.42) depends on the order r of the eigenmode. Therefore, the computational cost is proportional to the number of eigenmodes for which the sensitivity coeﬃcients are to be computed. Hence, this approach is not computationally eﬃcient for the case in which the eigenmodes corresponding to many eigenvalues are to be computed. Alternatively, the sensitivity coeﬃcients of the eigenmode can be expanded as follows using the eigenmodes based on the truncated modal method (Fox and Kapoor 1968): q ∑ ∂Φr = cj Φj (2.45) ∂Ai j=1 where q modes are assumed to be used, and c1 , . . . , cq are the coeﬃcients.

Design Sensitivity Analysis (1)

K1

FIGURE 2.5:

M0

(2)

71 M0

K2

A two-bar lumped mass structure.

From (2.36) and (2.37), we have Φ⊤ r KΦs = δrs Ωr , (r, s = 1, . . . , q)

(2.46)

By premultiplying Φ⊤ j on the both sides of (2.38), and using (2.37), (2.45), and (2.46), we obtain ( ) ∂K ∂Ωr ∂M cj (Ωj − Ωr ) = −Φ⊤ − M − Ω Φr (2.47) r j ∂Ai ∂Ai ∂Ai The coeﬃcients cj are found from (2.47) for j ̸= r. The coeﬃcient cr is obtained from (2.37), (2.41), and (2.45) as ∂M 1 Φr cr = − Φ⊤ 2 r ∂Ai

(2.48)

The computational cost is reduced by using only a few modes for expansion. However, as is seen from (2.48), if M does not depend on Ai , e.g., a lumped mass structure, then cr = 0 and the vector of sensitivity coeﬃcients of Φr is orthogonal to Φr . Therefore, higher modes are needed to accurately express the sensitivity coeﬃcients of the lower eigenmodes. Note also that this method cannot be used to compute the sensitivity coeﬃcients of an eigenmode associated with multiple eigenvalues, e.g., Ωr+1 = Ωr , because Ωj − Ωr in (2.47) vanishes for j = r + 1. Example 2.3 Consider a two-bar structure, as shown in Fig. 2.5, which has the lumped mass M0 at nodes 1 and 2. The structural mass of the bars is neglected for brevity. The extensional stiﬀnesses of bars 1 and 2 are denoted by K1 and K2 , respectively. Eq. (2.36) is written as ( ) ( ) K1 + K2 −K2 M0 0 Φr = Ω r Φr , (r = 1, 2) (2.49) −K2 K2 0 M0 The two eigenvalues are obtained as ( ) √ 1 Ω1 = K1 + 2K2 − K12 + 4K22 , 2M0 ( ) √ 1 K1 + 2K2 + K12 + 4K22 Ω2 = 2M0

(2.50)

72

Optimization of Finite Dimensional Structures

The sensitivity coeﬃcients are to be evaluated at the design (K1 , K2 ) = (3K0 , 2K0 ), where K0 is the speciﬁed value. Note that this relation deﬁnes only the values of the current design; i.e., it does not mean that K1 and K2 are linked by K0 . The eigenvalues and eigenmodes for this design are given as K0 6K0 , Ω2 = M0 M ( ) 0 ( ) 1 1 1 2 Φ1 = √ , Φ2 = √ 2 −1 5M0 5M0 Ω1 =

(2.51) (2.52)

Then the sensitivity coeﬃcient of Ω1 with respect to K1 is obtained from (2.39) as ( )( ) 1 ∂Ω1 1 1 0 1 = (1, 2) (2.53) = 0 0 2 ∂K1 5M0 5M0 which agrees with the result obtained by diﬀerentiation of Ω1 in (2.50). Next, we compute the sensitivity coeﬃcients of the lowest eigenmode Φ1 = (Φ1,1 , Φ1,2 )⊤ using (2.42). For the design (K1 , K2 ) = (3K0 , 2K0 ), the matrices K and K − Ω1 M are obtained as ( ) ( ) 5K0 −2K0 4K0 −2K0 K= , K − Ω1 M = (2.54) −2K0 2K0 −2K0 K0 As is seen from (2.54), the matrix K − Ω1 M is singular. From (2.42), we have √ M0 ∂Φ1,1 4K −2K − 0 0 5 1 ∂K1 √ √ − ∂Φ1,2 5M0 M0 = (2.55) −2K0 K −2 0 0 ∂K1 5 √ √ 0 ∂Ω1 M0 M0 − −2 0 ∂K1 5 5 from which we obtain ∂Φ1,1 4 ∂Φ1,2 2 ∂Ω1 1 √ √ =− , = , = ∂K1 ∂K1 5M0 25K0 5M0 25K0 5M0 ∂K1

(2.56)

Note that the sensitivity of the eigenvalue agrees with the result in (2.53). The vector b in (2.44) is given as √ M0 1 √ ∂Ω1 − 5 (2.57) b= − √ 5M0 ∂K1 M0 0 2 5 It is easily conﬁrmed that the sensitivity coeﬃcient derived from the orthogonality between b and Φ1 agrees with the result in (2.53).

Design Sensitivity Analysis The matrix B is found as

M0 − 5 √ M0 −2 5

73

√

4K0 −2K0 B= −2K0 K0

(2.58)

which has full row rank 2. Finally, we approximate the vector of the sensitivity coeﬃcient of Φ1 with respect to K1 at (K1 , K2 ) = (3K0 , 2K0 ) using Φ2 only as ∂Φ1 = c 2 Φ2 ∂K1 From (2.47), (2.51), and (2.52), we obtain ( )( ) ) 1 0 6K0 K0 1 ( 1 −2 1 + c2 = c2 0 0 2 5M0 M0 M0

(2.59)

(2.60)

which leads to

2 (2.61) 25K0 By incorporating c2 in (2.61) into (2.59), we obtain the same results as (2.56), because, in this example, the mass matrix is independent of K1 , and the sensitivity of Φ1 of the 2-degrees-of-freedom system can be exactly expressed by Φ2 . c2 =

2.3.2

Multiple eigenvalues

As discussed in Sec. 1.9, it is well known that the eigenvalues often become multiple as a result of optimization. In this case, the equations for sensitivity analysis of eigenvalues and eigenmodes presented in Sec. 2.3.1 cannot be used. Suppose the two lowest eigenvalues Ω1 and Ω2 have the same value at Ai = A∗i , where Ai is a representative design variable. Fig. 2.6(a) illustrates the variation of Ω1 with respect to Ai . Since Ω1 is deﬁned as the lowest eigenvalue, its variation is as plotted in the solid line in Fig. 2.6(a), and the two lowest eigenvalues coincide at Ai = A∗i , at which the derivatives of the eigenvalues are not continuous functions of the design variable Ai . Therefore, only the directional derivative or the subdiﬀerential can be deﬁned, as shown in Fig. 2.6(b), at Ai = A∗i ; see Appendix A.1.4 for deﬁnitions of the directional derivative and the subdiﬀerential. For a design satisfying Ω1 = Ω2 , the linear combination Ψ of the eigenmodes Ψ = a1 Φ1 + a2 Φ2

(2.62)

is also an eigenmode corresponding to the eigenvalue Ω1 (= Ω2 ), where a1 and a2 are the coeﬃcients satisfying a21 + a22 ̸= 0. Note that the orthogonality between Φ1 and Φ2 with respect to the mass matrix is satisﬁed.

74

Optimization of Finite Dimensional Structures

Ω2

Left directional derivative

Subdifferential of Wi

Eigenvalues Ωi

Double lowest eigenvalue

Ω2

Ω1

Subdifferential at A i = A *i

Right directional derivative

Ai

A *i

(a) Lowest eigenvalue

Ai

A *i

(b) Directional derivatives and subdiﬀerentials

FIGURE 2.6: Deﬁnition of directional derivatives and subdiﬀerentials of the lowest eigenvalue.

Deﬁne Crs as Crs =

Φ⊤ r

(

∂K ∂M − Ωr ∂Ai ∂Ai

) Φs , (r, s = 1, 2)

(2.63)

By premultiplying Ψ on both sides of (2.38) corresponding to r = 1, 2, and using (2.36), (2.62), and (2.63), we obtain ( )( ) ( ) ∂Ωr a1 C11 C12 a1 = (2.64) C21 C22 a2 ∂Ai a2 Therefore, the two directional derivatives are found as the solution of the eigenvalue problem (2.64) so that the derivatives and the coeﬃcient vector (a1 , a2 )⊤ have nonzero values (Haug and Choi 1982; Seyranian, Lund, and Olhoﬀ 1994). Example 2.4 Consider the four-bar truss, as shown in Fig. 1.16, that was used in Example 1.13 in Sec. 1.9. The extensional stiﬀnesses of the horizontal and vertical members are denoted by K1 and K2 , respectively. The truss has a nonstructural mass M0 , and the member mass is neglected. The eigenvalues and eigenvectors of this 2-degrees-of-freedom truss are deﬁned by ( ) ( ) 2K1 0 M0 0 Φr = Ωr Φr , (r = 1, 2) (2.65) 0 2K2 0 M0 The stiﬀness matrix K is diﬀerentiated with respect to K1 as ( ) ∂K 2 0 = 0 0 ∂K1

(2.66)

Design Sensitivity Analysis

75

If K1 = K2 = K0 , then the eigenvalues are duplicate as Ω1 = Ω2 =

2K0 M0

(2.67)

and the eigenmodes are found as ( ) ( ) 1 0 Φ1 = , Φ2 = 0 1

(2.68)

Then, from (2.63), we have C11 = 2, C12 = C21 = 0, C22 = 0

(2.69)

∂Ω1 = 2 or 0 ∂K1

(2.70)

which leads to

In fact, the increment of Ω1 is 0 for ∆K1 > 0, and 2∆K1 for ∆K1 < 0. Since the eigenmodes cannot be deﬁned uniquely for duplicate eigenvalues, suppose the two eigenmodes that are mutually orthogonal with respect to the mass matrix are deﬁned as (√ ) (√ ) 1 1 2 √ √2 Φ1 = , Φ2 = (2.71) 2 − 2 2 2 In this case, the coeﬃcients in (2.63) are obtained as C11 = 1, C12 = C21 = 1, C22 = 1

(2.72)

and (2.64) is written as ( )( ) ( ) ∂Ω1 a1 1 1 a1 = 1 1 a2 ∂K1 a2

(2.73)

which leads to the same result as (2.70).

Generally, for a symmetric structure, the matrix C = (Cij ) on the left-hand side of (2.64) is diagonal if the design variables are linked so that symmetry is preserved. In this case, the directional derivatives can be easily found from (2.39) for the simple eigenvalue using the symmetric (antisymmetric) eigenmodes corresponding to the multiple eigenvalues. For example, in Example 2.4, Φ1 = (1, 0)⊤ and Φ2 = (0, 1)⊤ in (2.39) lead to the sensitivity coeﬃcients 2 and 0, respectively, which are the same as those in (2.70).

76

Optimization of Finite Dimensional Structures 5 Node 2

4

2 1

6

Node 1 3

FIGURE 2.7:

2.4

Deﬁnition of local displacement numbers.

Linear buckling load

For structures such as frames with slender members and thin-walled shells, buckling is a very important performance measure in the design process. Although buckling is related to geometrically nonlinear behavior, the buckling load can be computed with good accuracy as a solution of a linear eigenvalue problem, if the deformation at buckling called prebuckling deformation is sufﬁciently small (see, e.g., Baˇzant and Cedolin (1991) and Koll´ar (1999) for details of buckling analysis of frames). Consider a truss subjected to the proportional loads ΛP0 , where Λ is the load factor, and P0 is the constant vector of load pattern. The load factor Λc at which the structure becomes unstable as Λ is increased from 0 is called the linear buckling load factor if the prebuckling deformation is negligibly small, and, hence, the stresses are proportional to the load factor. The displacement vector U0 against P0 is computed from KU0 = P0

(2.74)

The axial force N0i of member i of a truss corresponding to U0 is given by using the member displacement vector u0i with respect to the global coordinates as EAi ⊤ N0i = d u0i (2.75) Li i The reduction of stiﬀness due to compressive axial force is expressed by using the geometrical stiﬀness matrix. Suppose the local displacement numbers are deﬁned as shown in Fig. 2.7. Then the geometrical stiﬀness matrix kGi of member i with respect to the local coordinates is given for the load P0 as 0 0 0 0 0 0 0 1 0 0 −1 0 N0i 0 0 1 0 0 −1 kGi = (2.76) Li 0 0 0 0 0 0 0 −1 0 0 1 0 0 0 −1 0 0 1

Design Sensitivity Analysis

77

Note that N0i has a negative value in the compressive state. Because the elastic axial stiﬀness is suﬃciently larger than the geometrical stiﬀness, the eﬀect of axial force on the axial stiﬀness is usually neglected. However, we can incorporate its eﬀect by formulating kGi as 1 0 0 −1 0 0 0 1 0 0 −1 0 N0i 0 0 1 0 0 −1 kGi = (2.77) Li −1 0 0 1 0 0 0 −1 0 0 1 0 0 0 −1 0 0 1 which is compatible with the deﬁnition of Green’s strain and does not have any stiﬀness against rigid-body rotations. By assembling the member geometrical stiﬀness matrix to the total structure to formulate the n × n geometrical stiﬀness matrix KG (N0 ) as a function of N0 , the tangent stiﬀness matrix is given as K+ΛKG (N0 ), and the buckling load is obtained from the linear buckling analysis as (K + ΛKG )Φ = 0

(2.78)

In the following, the argument N0 is omitted for brevity. Note that KG is an indeﬁnite matrix, although the linear stiﬀness matrix K is positive deﬁnite, because the structure may buckle against the proportional loads in the opposite direction. However, we are interested in a positive buckling load factor for the load pattern P0 in the speciﬁed direction. Therefore, using the positive deﬁniteness of K in (2.78), the buckling mode Φ is normalized as Φ⊤ KG Φ = −1 (2.79) Note that Φ can alternatively be normalized with respect to K as Φ⊤ KΦ = 1; however, the normalization in (2.79) leads to a simple form of sensitivity coeﬃcient shown below. By diﬀerentiating (2.78) with respect to the design variable Ai , keeping in mind that KG depends on N0 , we have ) m ( ∑ ∂K ∂Φ ∂Λ ∂KG ∂N0j ∂Φ Φ+K + KG Φ + Λ Φ + ΛKG = 0 (2.80) ∂Ai ∂Ai ∂Ai ∂N ∂A ∂A 0j i i j=1 By premultiplying Φ on (2.80) and using (2.78), (2.79), and the symmetry of K and KG , we obtain ) ( m ∑ ∂K ∂Λ ∂KG ∂N0j = Φ⊤ Φ+Λ Φ⊤ Φ ∂Ai ∂Ai ∂N0j ∂Ai j=1

(2.81)

It is seen from (2.81) that the sensitivity coeﬃcients of the buckling mode are not needed for computing those of the linear buckling load factor, which

78

Optimization of Finite Dimensional Structures

is similar to the process of sensitivity analysis of eigenvalues of vibration in Sec. 2.3. However, the sensitivity coeﬃcients of the axial forces should be computed using (2.75), as shown in Sec. 2.2, and are needed for a statically indeterminate truss. For a statically determinate truss, by contrast, the sensitivity coeﬃcients of axial forces vanish, because the axial force is independent of the cross-sectional areas.

2.5 2.5.1

Transient responses Direct diﬀerentiation method

Consider a linearly damped ﬁnite-dimensional structure subjected to timevarying loads P(t). The displacement vector U(t) is also a function of time t. Let C denote the damping matrix. Diﬀerentiation with respect to t is indicated by a dot. The equation of motion is written as ¨ ˙ MU(t) + CU(t) + KU(t) = P(t)

(2.82)

with the initial conditions ˙ ˙0 U(0) = U0 , U(0) =U

(2.83)

˙ 0 are the speciﬁed initial displacement and velocity vectors, where U0 and U respectively. Suppose the dynamic transient response analysis has been completed to ˙ ¨ ﬁnd U(t), U(t), and U(t). By diﬀerentiating both sides of (2.82) with respect to Ai , we obtain ¨ ˙ ∂U ∂U ∂U M +C +K =Q (2.84) ∂Ai ∂Ai ∂Ai where Q=

∂P ∂M ¨ ∂C ˙ ∂K − U− U− U ∂Ai ∂Ai ∂Ai ∂Ai

(2.85)

It is assumed that the initial conditions are independent of the design variables as ˙ ∂U ∂U = 0, = 0 at t = 0 (2.86) ∂Ai ∂Ai It is seen from (2.84) and (2.86) that the sensitivity coeﬃcients are computed using conventional time-history analysis, where ∂U/∂Ai and Q are conceived as displacements and loads, respectively. Consider the case where the representative response quantity Fe(A), for which the sensitivity coeﬃcients are to be computed, is deﬁned by the integral

Design Sensitivity Analysis of a time-varying function R(U(A, t), A) as ∫ tf e F (A) = R(U(A, t), A)dt

79

(2.87)

where tf is the ﬁnal time. For example, if Fe (A) represents the L2 -norm of the displacement vector over the speciﬁed time period 0 ≤ t ≤ tf , R is given as √ R(U(A, t), A) = U⊤ (A, t)U(A, t) (2.88) Other norms, such as L1 -norm and L∞ -norm, can also be used (see Appendix A.4.4 for deﬁnitions of norms (distances)). The displacement Uj at the speciﬁed time t0 can also be formally deﬁned using the Dirac delta function δ(t0 ) as R(U(A, t), A) = Uj (A, t)δ(t0 ) (2.89) In the following, the arguments A, U, and t are omitted for brevity. The sensitivity coeﬃcients of the representative response Fe are computed from the following equation obtained by diﬀerentiating (2.87) with respect to Ai : ∫ tf n ∑ ∂ Fe (A) ∂R(U, A) A) ∂U ∂R(U, j dt = + (2.90) ∂Ai ∂Ai ∂U ∂A j i 0 j=1 This approach is classiﬁed as the direct diﬀerentiation method. Note that (2.90) is an integration of the known values, which can be computed by a simple algorithm of numerical integration, e.g., the trapezoidal rule, after ∂Uj /∂Ai (i = 1, . . . , n) are known. Therefore, the computation of the time history of the design sensitivity vector ∂U/∂Ai in (2.84) requires the most computational time; hence, the computational cost is proportional to the number of design variables.

2.5.2

Adjoint variable method

In the adjoint variable method, the adjoint variable vector λ(t) is deﬁned as follows in the time interval 0 ≤ t ≤ tf : ¨ ⊤ M − λ˙ ⊤ C + λ⊤ K = ∂R λ ∂U where ∂R = ∂U

(

∂R ∂R ,..., ∂U1 ∂Un

(2.91)

)⊤ (2.92)

By premultiplying λ⊤ on (2.84), and integrating with respect to t from 0 to tf , we obtain ) ( ∫ tf ˙ ¨ ∂U ∂U ∂U ⊤ +C +K − Q dt = 0 M (2.93) λ ∂Ai ∂Ai ∂Ai 0

80

Optimization of Finite Dimensional Structures

Integrating (2.93) by parts, we have t=tf t=tf ˙ ∂ U ∂U ⊤ ⊤ λ M +λ C ∂Ai ∂Ai t=0 t=0 ( ) ∫ tf ˙ ⊤ ⊤ ∂ U ∂U ∂U ⊤ + −λ˙ M − λ˙ C +λ K dt ∂Ai ∂Ai ∂Ai 0 ∫ tf = λ⊤ Qdt

(2.94)

Further integration leads to t=tf t=tf t=tf ˙ ⊤ ∂ U ∂U ∂U ⊤ ⊤ ˙ λ M +λ C −λ M ∂Ai ∂Ai t=0 ∂Ai t=0 t=0 ∫ tf ( ∫ tf ) ¨ ⊤ M − λ˙ ⊤ C + λ⊤ K ∂U dt = + λ λ⊤ Qdt ∂A i 0 0

(2.95)

The boundary terms in (2.95) vanish at t = 0, because the initial conditions for displacements and velocities are given as (2.86). The following conditions are to be satisﬁed for the adjoint variable vector so that the boundary terms in (2.95) vanish at t = tf : ˙ f) = 0 λ(tf ) = 0, λ(t

(2.96)

By incorporating (2.91) into (2.95), and using (2.96), we can obtain the following relation: ∫ 0

tf

n ∑ ∂R(U, A) ∂Uj j=1

∂Uj

∫

tf

λ⊤ Qdt

(2.97)

) ∂R(U, A) ⊤ + λ Q dt ∂Ai

(2.98)

∂Ai

dt = 0

Therefore, from (2.90) and (2.97), we have ∂ Fe(A) = ∂Ai

∫ 0

tf

(

Note that (2.98) is an integration of the known values, which can be computed using a simple algorithm of numerical integration. Therefore, the computation of the time history of the adjoint variable vector in (2.91) requires the most computational time; hence, the computational cost is proportional to the number of response quantities for which the sensitivity coeﬃcients are to be found.

Design Sensitivity Analysis

81

Λ

A i = A *i Λ = Λ∗

Ai = A *i + ∆Ai

Ui

FIGURE 2.8:

2.6

Geometrically nonlinear load-displacement relation.

Nonlinear responses

So far we considered the responses deﬁned by geometrically linear analysis, for which the linear relation (2.7) is satisﬁed between the loads and displacements. However, for ﬂexible structures, including shallow latticed domes and arch-type trusses in civil and architectural engineering, the eﬀect of geometrical nonlinearity should be incorporated in the strain-displacement relation for computation of responses against external loads. Fig. 2.8 illustrates the relation between the representative displacement component Uj and the load factor Λ for a proportional loading P = ΛP0 deﬁned with load pattern vector P0 . The equilibrium path plotted in the solid line may move to the dashed line due to a modiﬁcation of the design variable Ai , e.g., from A∗i to A∗i +∆Ai . The purpose of sensitivity analysis here is to compute the rate of change of displacements at the speciﬁed load factor Λ∗ . Consider an elastic conservative system that has the total potential energy Π(U(A), A) for the ﬁxed value of Λ, which is a function of U(A) and A. The equilibrium equations are obtained from the stationary condition of Π(U(A), A) as ∂Π(U(A), A) =0 (2.99) ∂U By diﬀerentiating (2.99) with respect to a design variable Ai , we can obtain the following equations for computing the sensitivity coeﬃcients of U with respect to Ai : ∂2Π ∂U + =0 (2.100) KT ∂Ai ∂Ai ∂U where KT is called the tangent stiﬀness matrix, for which the (i, j)-component KT,ij is given as ∂2Π (2.101) KT,ij = ∂Ui ∂Uj

82

Optimization of Finite Dimensional Structures ΛP

θ θ0 L

FIGURE 2.9:

L

A two-bar shallow truss. 0.6

Load factor Λ

0.4 0.2 0 −0.2 −0.4 −0.6 0

0.2

0.4

0.6

0.8

Displacement θ

FIGURE 2.10:

Equilibrium path of the two-bar shallow truss.

Hence, the tangent stiﬀness matrix is symmetric for a conservative system. For the case of proportional loading, Π(U(A), A) is deﬁned with the strain ∗ energy S(U(A), A) and the external work Λ∗ P⊤ 0 U at Λ = Λ as Π(U(A), A) = S(U(A), A) − Λ∗ P⊤ 0U

(2.102)

and (2.100) is reduced to KT

∂U ∂2S ∂P0 + = Λ∗ ∂Ai ∂Ai ∂U ∂Ai

(2.103)

which has a similar form as the geometrically linear case in (2.8). Indeed, if we use 1 (2.104) S(U(A), A) = U⊤ KU 2 then (2.103) reduces to (2.8). Example 2.5 Consider a two-bar shallow truss, as shown in Fig. 2.9, subjected to a vertical load ΛP , where the thin and thick lines represent the undeformed and deformed conﬁgurations, respectively.

Design Sensitivity Analysis

83

The cross-sectional areas and the elastic modulus are A and E, respectively, for both members. Therefore, the truss is symmetric with respect to the center vertical axis. The initial angle of each member to the horizontal axis is denoted by θ0 , and the rotation θ is taken as the generalized displacement. Since the deformed conﬁguration is also symmetric with respect to the center axis, the strain ε(θ) of both members is given as ε(θ) =

L/ cos(θ0 − θ) −1 L/ cos θ0

(2.105)

Then the total potential energy Π(θ, A) is written as Π(θ, A) = EAε2 (θ)

L − ΛP L[tan θ0 − tan(θ0 − θ)] cos θ0

(2.106)

By diﬀerentiating Π(θ, A) with respect to θ, we obtain the following equilibrium equation: 2EALε(θ)

sin(θ0 − θ) 1 − ΛP L 2 =0 cos2 θ0 cos (θ0 − θ)

(2.107)

From (2.107), we have the load-displacement relation ΛP = 2EA

sin(θ0 − θ)[cos(θ0 − θ) − cos θ0 ] cos(θ0 − θ)

(2.108)

which is plotted in Fig. 2.10 with P = 1, θ0 = π/6, E = 10, and A = 1. Note that Λ reaches the maximum at the limit point, as θ is increased, and attains 0 at θ = θ0 = π/6, when the two bars are colinear. Diﬀerentiation of (2.107) with respect to A at the ﬁxed value of Λ at Λ∗ leads to the following equation for computing the sensitivity coeﬃcient of θ with respect to A: { sin2 (θ0 − θ) cos(θ0 − θ) 2EAL − 2EALε(θ) 3 cos θ0 cos2 θ0 } (2.109) sin(θ0 − θ) ∂θ sin(θ0 − θ) ∗ + 2ELε(θ) 2 =0 + 2Λ P L 3 cos (θ0 − θ) ∂A cos (θ0 − θ)

2.7

Shape sensitivity analysis of trusses

So far, we considered cross-sectional areas as design variables for application to sizing optimization using a gradient-based approach. However, the nodal coordinates are to be modiﬁed in the process of shape optimization. Therefore,

84

Optimization of Finite Dimensional Structures

in this section, we derive the basic equations of shape sensitivity analysis of trusses with respect to nodal locations (Twu and Choi 1992). A threedimensional truss member, as shown in Fig. 2.3, is also used in this section. The 6 × 6 member stiﬀness matrix ki with respect to the global coordinates is given in (2.6). Because Li and the unit vector ti of member direction depend on the nodal coordinates, the sensitivity coeﬃcients of these values are needed to compute the shape sensitivity coeﬃcients of ki . Let (X1i , Y1i , Z1i ) and (X2i , Y2i , Z2i ) denote the global coordinates of nodes 1 and 2, respectively, of member i. Then Li and ti are given as √ Li = (X2i − X1i )2 + (Y2i − Y1i )2 + (Z2i − Z1i )2 (2.110) ti =

1 (X i − X1i , Y2i − Y1i , Z2i − Z1i )⊤ Li 2

(2.111)

By diﬀerentiating (2.110) with respect to X1i and X2i , respectively, we obtain ∂Li ∂Li 1 1 = − (X2i − X1i ), = (X i − X1i ) Li Li 2 ∂X1i ∂X2i

(2.112)

Hence, using (2.111) with the notation ti = (ti,1 , ti,2 , ti,3 )⊤ , simpler expressions ∂Li ∂Li = −ti,1 , = ti,1 (2.113) ∂X1i ∂X2i are obtained. The sensitivity coeﬃcients of ti,1 with respect to X1 and X2 are derived as 1 − t2i,1 1( ∂Li ) ∂ti,1 = = − − 1 − t i,1 Li Li ∂X1i ∂X1i 2 ( ) 1 − ti,1 1 ∂Li ∂ti,1 = 1 − ti,1 = i i Li Li ∂X2 ∂X2

(2.114)

The sensitivity coeﬃcients with respect to the Y - and Z-coordinates are derived in a similar manner. Then the sensitivity coeﬃcients of d are readily obtained from (2.4), and those of ki with respect to the nodal coordinates, e.g., X1i , are computed by diﬀerentiating (2.6) as Ai E ∂Li Ai E ∂d ⊤ Ai E ∂ki =− 2 di d⊤ d + d i + Li ∂X1i Li ∂X1i Li ∂X1i

(

∂d ∂X1i

)⊤ (2.115)

which is incorporated into (2.10) and (2.11) to compute the shape sensitivity coeﬃcients of the displacements. Therefore, the diﬀerence between design sensitivity and shape sensitivity exists only in the diﬀerentiation process of the local properties, and the formulations of the ﬁnal set of linear equations are the same.

Chapter 3 Topology Optimization of Trusses

In this chapter, various problem formulations and methods of topology optimization are presented for pin-jointed trusses. A brief introduction and historical review are given in Sec. 3.1. The traditional concept of the Michell truss is presented in Sec. 3.2. Some typical problem formulations and methodologies are summarized in Secs. 3.3 and 3.4. The problem under stress constraints is discussed for continuous variables in Sec. 3.5 and discrete variables in Sec. 3.6. A genetic algorithm is applied in Sec. 3.7 for the problem considering nodal costs. A random search method is presented in Sec. 3.8. An approach based on semideﬁnite programming for eigenvalue constraints is presented in Sec. 3.9. Finally, application of data mining is discussed in Sec. 3.10.

3.1

Introduction

Conﬁguration optimization of framed structures such as trusses and frames is categorized into two areas, i.e., optimization of nodal locations and optimization of connectivity of the nodes by the members. The former process is referred to as geometry optimization, and the latter is called topology optimization. Both processes usually include sizing optimization or traditional optimum design where the cross-sectional properties are optimized. Simultaneous optimization of topology and geometry may be referred to as layout optimization, or simply conﬁguration optimization; see Sec. 1.1 for details of the classiﬁcation and Sec. 1.10 for an illustrative example of geometry and topology optimization of a truss. The history of conﬁguration optimization of framed structures may be divided into three periods. Culmann (1875) seems to be the ﬁrst to have formulated the geometry optimization of trusses (Prager 1974b) to ﬁnd the fullystressed design. However, papers by Michell (1904) and Maxwell (1890) are usually cited as the ﬁrst work in the ﬁeld of layout optimization of trusses. Although the study by Michell is important in view of a theoretical basis (Hemp 1973), it can be applied to limited types of structures and constraints, and the so-called Michell truss is not practically acceptable because it has an inﬁnite number of members.

85

86

Optimization of Finite Dimensional Structures

In the 1960s and 1970s, when computers were available only to a limited number of researchers and engineers, some computational approaches under stress constraints were presented and small-scale toy problems were solved. In this period, many important theoretical results for numerical methods, including optimality criteria approaches, were presented, and diﬃculties in topology optimization were extensively studied. The theoretical works, e.g., the explicit optimality criteria approach and the method based on so-called grillage, are summarized in review papers by Kirsch (1989a) and books by Rozvany (1976, 1989, 1997). Recently, owing to the dramatic progress of computer technology, many numerical methods of topology optimization have been developed and applied to large-scale real-world structures. In this chapter, we concentrate on the numerical methods for topology optimization. Trusses are mainly considered here because rigidly jointed frames and pin-jointed trusses are modeled in similar formulations of the ﬁnite element method, and optimizing topologies of trusses is more diﬃcult than optimizing frames because of instability at the joints. In the widely used numerical approach to topology optimization, unnecessary members are removed from a highly connected ground structure while the nodal locations are ﬁxed (Kirsch 1989a; Topping 1992); see Sec. 3.3 for details. Many methods have been presented using this ground structure approach; therefore, ﬁnding practically acceptable solutions seems to be a matter of computational capacity. However, diﬃculties still exist for problems with the constraints on stresses, local buckling, and eigenvalues of vibration. Development of eﬃcient computational algorithms for optimizing large-scale trusses is also an important subject. Another diﬃculty in topology optimization is that the solution often turns out to be an unrealistic design due to the existence of unstable nodes, intersection of members, and the existence of extremely slender members. Nakamura and Ohsaki (1992) investigated the characteristics of optimal topologies under eigenvalue constraints and showed that local instability and multiplicity of eigenvalues lead to serious diﬃculties in ﬁnding optimal topologies. Kim, Jang, and Kim (2008) optimized topology of a frame with ﬂexible joints, where the joint stiﬀness is considered as a continuous design variable, while the sections of the beams are ﬁxed. Frame optimization with ﬂexible joints called 2-joints was also investigated for application to the framework of an automobile (Fredricson, Johansen, Klarbring, and Petersson 2003). Fredricson (2005) developed a material interpolation method, which is similar to continuum topology optimization, to optimize the ﬂexibility of joints. Various optimization results of frames are presented in Chap. 5. The methodologies of truss topology optimization can be applied to other ﬁelds of engineering, including ﬂow networks (Klarbring, Petersson, Torstenfelt, and Karlsson 2003; Evgrafov 2006).

Topology Optimization of Trusses

3.2

87

Michell truss

In his pioneering paper, Michell (1904) derived a condition for the optimal layout of a framework to have the least volume of material under stress constraints against static loads (Hemp 1973). Let σ t (> 0) and σ c (< 0) denote the bounds of stress in tension and in compression, respectively. Using σ0 = (σ t − σ c )/2 and a positive constant ε0 , the condition for optimality by Michell may be stated using the current terminologies as follows: • A truss has minimum material volume, if – the equilibrium conditions are satisﬁed against the speciﬁed static loads, – the stress of each member is equal to the upper bound σ t or the lower bound σ c , and – there exists a deformation compatible with the strains that are equal to σ0 ε0 /σ t for members in tension and σ0 ε0 /σ c for those in compression. Accordingly, the bars of the optimal pin-jointed framework are fully stressed and arranged in the directions of the principal strains of the displacement ﬁeld. However, these conditions do not uniquely determine the displacement ﬁeld. Therefore, the optimal layout should be searched using a trial-and-error process. The structure obtained this way is called the Michell structure or the Michell truss. Hegemier and Prager (1969) derived explicit optimality conditions for the Michell truss and presented an approach without trial and error. They also showed that the truss with maximum stiﬀness is fully stressed and extended the concept of the Michell truss to optimization under frequency constraints and for creep problems. Schmidt (1962) extended Michell’s theory to a statically determinate structure under multiple loading conditions. He showed that the condition of fully-stressed design is a necessary, but not suﬃcient, condition of optimality. Exact analytical benchmark solutions were presented by Lewi´ nski and Rozvany (2007). It is derived from the condition of the Michell truss that the inﬁnitely dense members exist in the two perpendicular directions of the principal strains of the displacement ﬁeld. To overcome the practical diﬃculty due to the existence of too many members, Prager (1974b) introduced the cost of node, or connection. Assuming that the solution is fully stressed, the cost coeﬃcient for the cross-sectional area Ai is increased from the member length Li by a certain speciﬁed amount b0 representing the cost of the connections; i.e., the total structural cost V is formulated as m ∑ V = Ai (Li + b0 ) (3.1) i=1

88

Optimization of Finite Dimensional Structures

P FIGURE 3.1:

A highly connected ground structure.

P FIGURE 3.2:

P

Illustration of topology optimization.

where m is the number of members. In this case, the conditions for optimality are obtained as Li |εi | = (Li + b0 )ε∗ for Ai > 0, (i = 1, . . . , m)

(3.2)

where ε∗ is a positive constant, and εi represents the strain of the ith member. Note that εi may represent the strain rate for a plastic limit design problem. This way, the number of members is successfully reduced to obtain a practically acceptable layout.

3.3

Topology optimization problem

Consider a problem of ﬁnding optimal locations of members, or connectivity of the nodes by members, from a set of existable nodes and members under

Topology Optimization of Trusses

89

constraints on the responses against speciﬁed static loads. The structure with all the candidate nodes and members is called the ground structure, which includes many nodes and members, as illustrated in Fig. 3.1; preferably a member should exist between every pair of nodes including supports. The approach that removes unnecessary members from the ground structure by optimization is called the ground structure approach, and the paper by Dorn, Gomory, and Greenberg (1964) is usually cited as the ﬁrst work on this approach. Fig. 3.2 illustrates the process of topology optimization of a plane truss with six nodes including two supports, where the dotted lines indicate the members to be removed after optimization. There have been extensive studies on this approach since the 1960s, e.g., Dobbs and Felton (1969); see, e.g., the review articles by Kirsch (1989a), Topping (1984), and the book by Rozvany (1997). Let Ai and Li denote the cross-sectional area and the length of the ith member, respectively. The number of members in the initial ground structure is denoted by m. Only inequality constraints Hj ≤ 0 (j = 1, · · · , nI ) are considered, for simplicity, because the design requirements in various ﬁelds of engineering are generally given with inequality. If the responses represented by displacements and stresses are considered to be implicit functions of the continuous design variables A = (A1 , . . . , Am )⊤ , then Hj is conceived as a function of A only, as presented in Sec. 1.3. Therefore, the problem of minimizing the total structural volume V (A) turns out to be a nonlinear programming (NLP) problem as

Minimize

V (A) =

m ∑

Ai Li

(3.3a)

i=1

subject to Hj (A) ≤ 0, (j = 1, · · · , nI ) Ai ≥ 0, (i = 1, · · · , m)

(3.3b) (3.3c)

where the upper bound for Ai is not given, for simplicity. The members with null cross-sectional areas are removed after optimization to obtain the optimal topology. Although the cross-sectional areas are assumed to be continuous variables in the conventional ground structure approach, the topology optimization problem under stress constraints is virtually a combinatorial optimization problem with 0–1 variables deﬁning the existence of members even for the case where Ai is a continuous variable, as demonstrated in Sec. 3.5. If the state variables deﬁned by nodal displacements are regarded as independent continuous variables, the problem becomes a mixed integer linear or nonlinear programming problem (Floudas 1995). Suppose Ai is selected from a discrete set of available sections. Then, the truss topology optimization problem for minimizing the total structural

90

Optimization of Finite Dimensional Structures

volume is formally written as Minimize

V (A) =

m ∑

Ai Li

(3.4a)

i=1

subject to Hj (A) ≤ 0, (j = 1, · · · , nI ) Ai ∈ Ai , (i = 1, · · · , m)

(3.4b) (3.4c)

where Ai represents the feasible set of Ai ; e.g., Ai = {0, A0 } if the crosssectional area of an existing member can take only the speciﬁed value A0 , and Ai = {0, A01 , . . . , A0r } if Ai can have one of the pre-deﬁned positive values A01 , . . . , A0r for an existing member and 0 for a nonexistent member.

3.4

Optimization methods

For the case in which constraints on displacements or compliance against static loads is considered, the topology optimization problem is formulated as an NLP problem based on the ground structure approach, and optimal solutions may be easily found by using an appropriate NLP algorithm or an optimality criteria approach in Sec. 1.7. Among many methods of NLP, sequential quadratic programming and the method of the moving asymptote are often used for structural optimization (see Appendix A.2.2 for details of NLP). Recently, new algorithms, including the interior-point method (Ben-tal and Nemirovski 1994; Koˇcvara 1997) and semideﬁnite programming (Ohsaki, Fujisawa, Katoh, and Kanno 1999), have been shown to be very eﬀective for special problems under constraints on eigenvalues and/or compliance. In the process of removing unnecessary members, the truss may become unstable due to the existence of a node without any connecting member or the existence of a hinge that connects only two colinear members. A lower bound ALi is therefore given as follows to prevent instability: Ai ≥ ALi , (i = 1, · · · , m)

(3.5)

Note that ALi has a very small positive value, and the member with Ai = ALi should be removed from the optimal solution; i.e., ALi is introduced to avoid numerical instability in the optimization process, and not to avoid the unstable optimal solution. Petersson (2001) showed that the optimal solution and the corresponding displacements for the limit ALi → 0 converge to the values for ALi = 0 if the strain energy of the solution with ALi = 0 is bounded. Svanberg and Werme (2009) discussed the validity of assigning small lower bounds for a problem with 0–1 variables. The possibility of obtaining an unstable solution may be reduced by considering multiple loading conditions, as demonstrated in Sec. 1.5. The method

Topology Optimization of Trusses

91

based on boundary cycle to restrict the feasible solution to the truss consisting of stable units only may also be useful (Kaveh 1986; Nakanishi and Nakagiri 1996a, 1996b). Including constraints on the shape and topology of the triangular unit is very eﬀective to obtain practically acceptable designs (Nakanishi and Nakagiri 1997; Ohsaki and Kato 1999; Kawamura, Ohmori, and Koto 2002). Ohsaki and Katoh (2005) assigned a lower bound for the number of members connected to an existing node to ensure stability (see Sec. 3.5.6 for details). Ohsaki and Watada (2008) and Watada and Ohsaki (2009b, 2009c) presented a mixed integer programming approach to select the layout of each unit of a regular truss from the predeﬁned standard layouts; see Sec. 3.6 for details. The stable topologies of bar-joint structures are extensively investigated in the ﬁeld of structural rigidity in applied mathematics utilizing graph theoretical approaches (Kaveh 1991; Graver, Servatius, and Servatius 1993; Kaveh 2004; Avis, Katoh, Ohsaki, Streinu, and Tanigawa 2007). Although an optimal solution is found by simply applying an NLP algorithm, it is very diﬃcult to ﬁnd the global optimal solution even for a simple problem with stress and/or displacement constraints, as demonstrated in Sec. 3.5. Minimization of the compliance under a single loading condition without an upper bound for the cross-sectional area is considered as a min-max problem for minimizing the compliance with respect to the design variable, simultaneously maximizing the total potential energy with respect to the displacements (Ben-Tal and Bendsøe 1993; Bendsøe, Ben-Tal, and Zowe 1994). A strictly convex dual problem is then formulated and is eﬃciently solved by a primal-dual approach (Beckers and Fleury 1997) (see Sec. 1.8.3 for several reformulations of the truss optimization problem under compliance constraints). Heuristic approaches, e.g., genetic algorithm (GA), can also be applied to truss topology optimization problems (Jenkins 1991; Ohsaki 1995); see Sec. 3.7. Recently, some evolutionary approaches have been presented (Dasgupta and Michalewicz 1997; Hajela 1997; Kwan 1998; Yang, Xie, Steven, and Querin 1999a; Lagaros, Papadrakakis, and Kokossalakis 2002; Greiner, Emperador, and Winter 2004; Ebenau, Rottsch¨afer, and Thierauf 2005; Kicinger, Arciszewski, and de Jong 2005). The term evolutionary is very confusing because it is used with several diﬀerent meanings. A GA with evolving parameters may be called evolutionary, or even a simple GA may be an evolutionary approach (Kwan 1998). The algorithm with local rules that is similar to the optimality criteria approaches and many variants of growing processes are also evolutionary. Simulated annealing (SA) is also useful for topology optimization (Topping, Khan, and Leite 1996; Tagawa and Ohsaki 1999). The advantage of using SA is that both continuous and discrete variables can be included without any diﬃculty. Tabu search may also be very eﬀective for truss topology optimization (Bennage and Dhingra 1995). Luh and Lin (2008) used the ant system for truss topology optimization. Because the computational cost of analysis

92

Optimization of Finite Dimensional Structures

Intersection of members FIGURE 3.3:

Very thin member

Unstable node

Infeasible topologies of trusses.

is very large for these heuristic methods, an eﬃcient method of reanalysis presented in Sec. 3.8.2 may be utilized for reducing the computational cost (Kirsch 1993, 1995; Kirsch and Liu 1995). The ground structure approach is regarded as the standard procedure for topology optimization, although there are many diﬃculties, which are summarized as follows: 1. Too many members and nodes are needed in the initial ground structure, because only removal of members is possible, and addition of members and nodes is very diﬃcult without resort to a heuristic approach. 2. The optimal topology strongly depends on the initial design, and an inﬁnite number of nodes and members is needed if the nodal locations are also to be optimized. 3. Unrealistic optimal solutions with very long members, intersection and/or overlap of members, etc., are often obtained. 4. The truss becomes unstable due to the existence of a node connected by two colinear members only, if too many members are removed. The unrealistic or infeasible topologies are illustrated in Fig. 3.3, including intersecting members, very thin members, and unstable nodes. Note that intersection of members is sometimes allowed, e.g., for a truss or a frame with moderately thin braces; see Sec 5.1.4. The nodal locations can also be considered as variables in order to optimize the geometry of a truss (Ohsaki and Nishiwaki 2005). This process is sometimes referred to as the extended ground structure approach.

Topology Optimization of Trusses

3.5 3.5.1

93

Stress constraints Introduction

One of the main diﬃculties in topology optimization under stress constraints is that the constraints need not be satisﬁed by the member with vanishing cross-sectional area; i.e., the constraint obviously does not exist for a member that does not exist, as noted by Sved and Ginos (1968), Sheu and Schmit (1972), Nagtegaal (1973), and Kirsch (1989a, 1990). Therefore, the problem becomes discrete in the formulation of constraints with a 0–1 variable indicating existence/nonexistence of a member, and it is very diﬃcult to ﬁnd the global optimal solution even for a simple truss topology optimization problem with a few members. The constraint of this type is called the design dependent constraint. If we consider a process of continuously decreasing the cross-sectional area of a member from a ﬁnite positive value to 0, the constraint suddenly disappears at the ﬁnal state with null cross-sectional area. As a result, the feasible region is not convex in general; hence, the optimal solution is often located at a cusp or an isolated point of the feasible region, as illustrated in Sec. 3.5.3. The optimal topology under stress constraints considering a single loading condition can be derived as a solution of a linear programming (LP) problem with member forces as design variables (Dorn, Gomory, and Greenberg 1964; Ringertz 1985). The solution is globally optimal if the optimal truss is statically determinate (Sved 1954). For the case in which the optimal topology is unstable due to the existence of colinear members connected by a hinge, as shown in Fig. 3.3, some members should be added or the unstable joints should be ﬁxed. To overcome the diﬃculty due to discontinuity in the problem formulation, several relaxation approaches as well as branch-and-bound-type iterative methods have been presented. Ringertz (1985) presented a method for problems with stress and displacement constraints, where a compatible set of strains and displacements is ﬁrst calculated for speciﬁed cross-sectional areas. An NLP problem is then solved under stress constraints while ﬁxing the displacements. Ringertz (1986) proposed an approach to obtaining the lowerbound solution by solving an NLP problem under displacement constraints only. The stress constraints are successively given for members that violate the constraints to obtain an upper-bound solution. The eﬃciency of the solution may be evaluated through comparison of the objective value with a lower-bound value that can be found by solving an LP problem neglecting the compatibility conditions (Kirsch 1989b). A relaxation method has been presented by Cheng (1995) and Cheng and Guo (1997) for obtaining a good approximate solution. In their approach, called ε-relaxed approach, the stress constraint is relaxed for a member with

94

Optimization of Finite Dimensional Structures

a small cross-sectional area. It is very diﬃcult, however, to determine an appropriate value of the parameter for relaxing the constraints and to assign the initial solution that can reach the global optimal solution, as noted by Stolpe and Svanberg (2001b). The convergence properties of this approach were discussed by Bruggi (2008). Guo and Cheng (2000) extended their method to incorporate an extrapolation approach (Nakamura and Ohsaki 1992; Ohsaki and Nakamura 1996). Duysinx and Sigmund (1998) presented a relaxation approach to a continuum topology optimization problem. Evgrafov and Patriksson (2003) presented a stochastic relaxation approach to stress constrained problems. Stolpe and Svanberg (2003b) showed that a stress constrained problem can be solved by a traditional NLP approach. Sui, Du, and Guo (2006) presented a method for topology optimization of frames using diﬀerent ﬁlter functions for weight, stiﬀness, and stress bound of the frame element.

3.5.2

Governing equations

Consider an elastic truss subjected to multiple static loading conditions. The problem is to obtain an optimal topology and the optimal member crosssectional areas that minimize the total structural volume under constraints on stresses of members, where the conventional ground structure approach is used. Let Pk denote the vector of the kth set of nodal loads. The vector of axial k ⊤ ) , where m is forces corresponding to Pk is denoted by Nk = (N1k , . . . , Nm the number of members in the ground structure. The equilibrium equations are given as DNk = Pk , (k = 1, . . . , nP ) (3.6) where nP is the number of loading conditions, and the n × m matrix D is called the equilibrium matrix with n being the number of degrees of freedom. Let Uk denote the vector of nodal displacements against Pk . The elongation of the ith member corresponding to Uk is denoted by δik . The compatibility condition between Uk and δik is written as k k P d⊤ i U = δi , (i = 1, . . . , m; k = 1, . . . , n )

(3.7)

where di is the ith column of D. Let εki and σik denote the strain and stress, respectively, of the ith member for the kth loading condition. Then, the strain, stress, and axial force of the ith member against the kth loading condition are obtained from Uk as εki =

δik , σik = Eεki , Nik = Ai σik Li

(3.8)

where Ai and Li are the cross-sectional area and length of the ith member, respectively, and E is the elastic modulus. Eqs. (3.6)–(3.8) are combined into

Topology Optimization of Trusses

95

2 1

π /4 π /4

3

αk Pk FIGURE 3.4:

A three-bar truss subjected to multiple loading conditions.

the following stiﬀness equation: KUk = Pk , (k = 1, . . . , nP )

(3.9)

where K is the stiﬀness matrix.

3.5.3

Discontinuity in stress constraint

The diﬃculties in topology optimization under stress constraints considering multiple loading conditions were ﬁrst discussed by Sved and Ginos (1968) using a simple three-bar truss as follows: Example 3.1 Consider a three-bar truss as shown in Fig. 3.4. The three loading conditions are deﬁned as (P k , αk ) = (40, π/4), (30, π/2), (20, 3π/4), where P k and αk denote the magnitude and direction of the kth load, respectively, as deﬁned in Fig. 3.4. The lower and upper bounds (σiL , σiU ) for the stress are (−5, 5), (−20, 20), and (−5, 5), for members 1, 2, and 3, respectively. Suppose the length of member 2 is equal to 1, for simplicity. The objective function is the total structural volume that is given as √ V = A2 + 2(A1 + A3 ) (3.10) The optimal solution that minimizes V under stress constraints for three loading conditions is A1 = 8.0, A2 = 1.5, and A3 = 0, where V = 12.812. The stresses of three members are (5, 0, −5), (0, 20, 20), and (−2.5, 18.856, 21.356) for the three loading conditions. Note that the stress of member 3 for P 3 is 21.356, which is greater than the upper bound 5. The stress constraint, however, need not be satisﬁed, because A3 = 0 and member 3 does not exist. Note that the stress of member 3 can be computed from the nodal displacements as (3.7) and (3.8).

96

Optimization of Finite Dimensional Structures 6

4 A3

Feasible region 2 Optimal solution B

A

0 0

2

4

6

8

10

12

A2

FIGURE 3.5:

Feasible region of the three-bar truss.

If we assign the stress constraint even for the nonexistent member, the optimal solution is found to be A1 = 7.099, A2 = 1.849, and A3 = 2.897, where V = 15.986, which is greater than the objective value 12.812 of the global optimal solution. This result suggests that the optimal solution exists at a singular point ‘A’ in the feasible region, which is plotted in Fig. 3.5 in the A2 –A3 plane for A1 = 8.0. Note that the feasible region includes the line AB in addition to the shaded region. Incorporating the discontinuity into the constraints, the topology optimization problem with stress constraints is formulated as m ∑

Minimize

V =

subject to

i=1 σiL ≤ σik

Ai Li

(3.11a)

≤ σiU for Ai > 0,

(i = 1, . . . , m; k = 1, . . . , nP ) Ai ≥ 0, (i = 1, . . . , m)

(3.11b) (3.11c)

The stress may be calculated from axial force Nik as σik = Nik /Ai ; however, this relation cannot be used for a removed member because both Nik and Ai vanish. On the other hand, the strain εki can be calculated from the elongation δik , which is easily found from the displacements of the nodes connected by the removed member, if they exist, and the stress is calculated using the relations (3.8) (Cheng and Jiang 1992) as we have done in Example 3.1. Therefore, there is no discontinuity in σik itself at Ai = 0 if the truss after removal of the ith member is stable, and if the stress is obtained from the strain, not from the axial force; i.e., only the deﬁnition of the stress constraints (3.11b) is discontinuous. A constraint of this type is said to be a design dependent constraint or a vanishing constraint (Rozvany 2001; Izmailov and

Topology Optimization of Trusses

97

L (1)

(3)

1

P 2 4 y

FIGURE 3.6:

L

5 3

(2)

(4)

x

P

A ﬁve-bar truss.

Solodov 2008). The optimal topology with violating the stress constraint in the vanishing members is called singular optimal topology. Constraint (3.11b) suggests that the stress constraints should be relaxed in the vicinity of Ai = 0 if the optimal topologies are to be found by the ground structure approach. Cheng and Guo (1997) presented the ε-relaxed approach that reformulates the stress constraint as (σiL − σik )Ai ≤ e

(3.12a)

(σik − σiU )Ai ≤ e Ai ≥ e2

(3.12b) (3.12c)

where the parameter e has a suﬃciently small positive value. The optimal topology may be found by successively reducing e to 0. Example 3.2 As a small example, consider a ﬁve-bar truss, as shown in Fig. 3.6, that was solved by Cheng and Guo (1997). The units for force and length are omitted for brevity. Two loading conditions are considered, where the loads P = (Px , Py )⊤ = (5, −50)⊤ are applied at nodes 3 and 4, respectively. The parameters are E = 1.0 and e = 0.0001. The upper and lower bounds for stresses are ±20 for member 2 and ±5 for the remaining members. The optimization software package IDESIGN Ver. 3.5 (Arora and Tseng 1987), which utilizes sequential quadratic programming, is used. If the initial design is selected as (A1 , . . . , A5 ) = (5.0, 8.0, 15.0, 1.0, 20.0), then the global optimal topology is successfully found by the ε-relaxed approach with (A1 , . . . , A5 ) = (1.0000, 2.5000, 10.0000, 0.0, 14.1421) and V = 35.520. Note that the truss consists of four members, and the stress of the nonexistent member 4 for the load at node 3 is −15.0, which violates the constraint.

98

Optimization of Finite Dimensional Structures

However, if the initial solution is given as Ai = 10.0 for all the members, a non-optimal solution (A1 , . . . , A5 ) = (7.2194, 4.7319, 4.5318, 8.7956, 7.8310) with V = 39.986 is reached. In order to investigate the convergence property, optimal solutions are found starting with cross-sectional areas with uniform random distribution between 0 and 20.0. Among the ten trials, the global optimal solution was found twice. The numbers of trials that lead to the non-optimal solution with ﬁve bars, four bars without members 3 and 5, respectively, are 4, 1, and 3.

To prevent convergence to a local optimal solution, it is natural to use a continuation method, where the parameter e in (3.12) is successively reduced to 0, while the optimal solution for a certain value of e is chosen as the initial solution for optimization corresponding to a slightly reduced value of e. However, even using this approach, convergence to the global optimal solution is not guaranteed (Petersson 2001; Stolpe and Svanberg 2001b), because the trajectory of the solution with respect to the parameters e may be discontinuous near the global optimal solution.

3.5.4

Discontinuity due to member buckling

For pin-jointed trusses, the lower-bound stress σiL (< 0) should be deﬁned by the Euler buckling stress if local buckling is considered (Cheng and Guo 1997). It is known that the optimal topology strongly depends on the values of stress bounds in tension and compression especially for bridge-type structures (Achtziger 1996; Oberndorfer, Achtziger, and H¨ornlein 1996). As a simple example, consider a four-bar truss, as shown in Fig. 3.7, subjected to a compressive load to the horizontal members 1 and 2. The lengths of members 1 and 2 are L, and the radius of gyration is r irrespective of the cross-sectional area; i.e., we assume a sandwich section, a rectangular section with constant height, or a pipe section with an appropriate relation between the radius and thickness. If we assume P > 0, members 1 and 2 are in a compressive state, and the lower-bound stress is deﬁned with the Euler buckling stress −π 2 E/(r/L)2 , of which the absolute value is smaller than the yield stress if the slenderness ratio is larger than the critical slenderness ratio. This dependence of the lower-bound stress on the member length leads to more diﬃculties in topology optimization of pin-jointed trusses. In the practical design process, the lower-bound stress should be deﬁned by dividing the Euler buckling stress by a safety factor, which is not considered here for simplicity. Since the stresses of vertical members 3 and 4 are 0 for the truss in Fig. 3.7, these members are removed as the result of optimization. Hence, the optimal solution has two colinear members 1 and 2 connected by pin joint 2, and the optimal truss is unstable. The unstable truss can be stabilized by ﬁxing pin joint 2. Consequently, a long member with length 2L emerges, and the

Topology Optimization of Trusses L

99

L

(4) (1)

1

(2)

3 2 4

(3)

P

H H

(5)

FIGURE 3.7:

A four-bar truss under a compressive load.

lower-bound stress should be replaced by −π 2 E/(r/2L)2 , which is 1/4 of the initially assigned value. Therefore, the stress constraint will be violated by the optimal solution that satisﬁes the stress constraints with equality in the short two members 1 and 2. One possible strategy to avoid the existence of unstable optimal topology may be to include constraints on global buckling or imperfection of nodal location in the problem formulation (Zhou 1996) (see Sec. 3.9.4 for topology optimization under linear buckling constraints). In this case, unrealistically very slender members exist in the optimal topology, as pointed out by Nakamura and Ohsaki (1992) for a problem with frequency constraints; see Sec. 3.9.2. For the four-bar truss in Fig. 3.7, members 3 and 4 will have very small crosssectional areas to add vertical stiﬀness at node 2. The solution with thin bracing members obtained this way usually has a smaller objective value than the solution with a long member after ﬁxing the unstable nodes and increasing the cross-sectional areas to satisfy the local buckling constraints. However, in some cases the solution with thin bracing members has a larger objective value than that with long members in compression (Rozvany 1996). For example, suppose H ≫ L in Fig. 3.7, although this is a very unrealistic e1 and A e2 denote the optimal cross-sectional areas before situation. Let A e e2 = A e = −P/σ L , where σ L = −π 2 E/(r/L)2 and ﬁxing node 2; i.e., A1 = A e3 = A e4 = 0. After ﬁxing node 2, the cross-sectional areas of members 1 and A e to satisfy the Euler buckling constraint; hence 2 should be increased to 4A e If the global buckling the total structural volume increases to Ve = 8LA. constraint is introduced in the optimization problem of the four-bar truss with an appropriately small lower-bound linear buckling load, members 3 and 4 have a small cross-sectional area denoted by AL , and the cross-sectional areas e at the optimal solution. Then the total structural of members 1 and 2 are A e + 2HAL , which is larger than Ve if H > 3LA/A e L . The value of volume is 2LA L A may be given by incorporating the buckling constraint also for the bracing members, as discussed by Takagi and Ohsaki (2004) for the design of columntype frames with lateral braces. Achtziger (1999a, 1999b) deﬁned chain as a

100

Optimization of Finite Dimensional Structures P (2) L

π/4 L

brace column

(3)

(1)

FIGURE 3.8:

Cantilever column supported by a brace.

set of colinear members without connecting members at the unstable nodes, and assigned the Euler buckling constraints using the buckling length of the active chain. He also discussed the possibility of obtaining a stable truss by introducing constraints on the slenderness ratio. Guo, Cheng, and Yamazaki (2001) used a quadratic extended interior penalty function (Haftka and Starnes 1976) for approximating the local buckling constraints of members with small cross-sectional areas, where the section is assumed to be a circular solid. Evgrafov (2005) noted, for the problem of minimizing the total structural volume under a linear buckling constraint, that the solution with a small lower-bound cross-sectional area has far larger objective value than that with the vanishing lower bound, because the existence of the nodes connected by slender members only may have local (nodal) buckling under very small loads. Note that local instability can be avoided if rigidly jointed frame elements are used. Ohsaki (1997b) used frame elements for conﬁguration optimization; see Chap. 4. A semi-rigid connection can also be used for preventing local instability (Ramrakhyami, Frecker, and Lesieutre 2009). Instability of the optimal truss is closely related to nonuniqueness of the displacement (Koˇcvara and Outrata 2006), which may lead to the nonuniqueness of optimal crosssectional areas. Therefore, a unique solution can be found by minimizing a strictly convex function with respect to the cross-sectional areas and displacements among the set of nonunique solutions. For example, a physical displacement can be obtained by minimizing the norm of transverse displacements of the members related to Green’s strain. Uniqueness of the optimal solutions of building frames is discussed in Sec. 5.2. Example 3.3 Discontinuity due to member buckling also exists in the topology optimization problem of a braced frame. Consider, for example, a cantilever column, as shown in Fig. 3.8, subjected to a lateral load P and supported by a brace

Topology Optimization of Trusses

101

modeled by a truss member (Hagish*ta and Ohsaki 2007). Let Ac and Ic denote the cross-sectional area and the second moment of inertia of the column, respectively. The cross-sectional area of the brace is denoted by Ab . The column and brace have the same length L. The stress σb of the brace is derived as √ 2P L3 Ac (3.13) σb = − 6Ic Ac + 3Ic Ab + Ab Ac L2 First, we assume that the lower-bound stress of the brace is independent of eb of Ab at which σb is equal to the lower Ab . Then, from (3.13), the value A L bound σ (< 0) is derived as √ Ac ( 2P L3 Ac + 6σ L Ic ) e Ab = − σ L (3Ic + Ac L2 )

(3.14)

Consider a problem of minimizing the structural volume of the brace under constraint σb ≥ σ L of the stress of the brace, while ﬁxing the cross-section of the column; i.e., the cross-sectional area of the brace is simply minimized in eb cannot have a positive value, this small example. It is seen from (3.14) that A i.e., the stress√constraint is satisﬁed for Ab ≥ 0 if Ic has a suﬃciently large value so that 2P L3 Ac + 6σ L Ic < 0 is satisﬁed with σ L < 0. Therefore, the √ brace is not needed if the column is suﬃciently stiﬀ. In contrast, if 2P L3 Ac + eb satisfying (3.14), and the 6σ L Ic > 0, then there exists a positive value A e stress constraint is violated for 0 < Ab < Ab . However, no stress constraint is needed for the brace if Ab = 0. Therefore, the feasible region is discontinuous, and the optimal solution Ab = 0 exists at an isolated point of the feasible region. Next, we consider the member buckling constraint and assume that the Euler buckling stress is proportional to Ab , which holds, for example, for a solid rectangular section with variable height and constant width. In this case, σ L is deﬁned to be proportional to Ab with a positive coeﬃcient α as σ L = −αAb , and (3.13) with σb = σ L reduces to α(3Ic + Ac L2 )A2b + 6αIc Ac Ab −

√

2P L3 Ac = 0

(3.15)

Therefore, there always exists a positive value Ab satisfying the stress constraint with equality; hence, the feasible region is disjoint and the optimal solution exists at the isolated point Ab = 0.

3.5.5 3.5.5.1

Mathematical programming approach Reformulation of the problem

As we have seen in Example 3.3, a singular optimal topology with violated stress constraints by vanishing members cannot be generally found by solving the original NLP problem (3.11) using a gradient-based NLP algorithm.

102

Optimization of Finite Dimensional Structures

Therefore, in order to resolve the discontinuity of the problem, several methods have been presented for reformulation of the problem. Equivalence among the problems under constraints on compliance, stress, and plastic limit load has been extensively studied since the 1970s (Hemp 1973). A problem with a single loading condition can be written in terms of the member forces and the cross-sectional areas as (Kirsch 1989b; Muralidhar and Rao 1997; Stolpe and Svanberg 2003b) Minimize

V (A) =

m ∑

Ai Li

(3.16a)

i=1

subject to DN = P

(3.16b)

≤ Ni ≤ (i = 1, . . . , m) Ai ≥ 0, (i = 1, . . . , m) Ai σiL

Ai σiU ,

(3.16c) (3.16d)

which is an LP problem with variables A and N, and the superscript indicating load number is omitted. The redundant forces are often used as independent variables in Problem (3.16) of a statically indeterminate truss to reduce the number of variables using the equilibrium conditions (3.16b), and to investigate the properties of the optimal solutions. Sedaghati and Esmailzadeh (2003) used singular value decomposition (see Appendix A.1.3) for extracting the independent force components. Note that this approach is widely used for tension structures (Pellegrino 1993; Ohsaki, Zhang, and Ohishi 2008). However, we use the force vector of all members for a simple presentation of the formulations. Because the existence of displacements compatible to A and N is not assured in Problem (3.16), the solution to this problem gives a lower bound for the objective value of the original NLP problem with compatibility conditions. However, there always exists a displacement vector compatible to A and N for the solution of Problem (3.16), because the solution of the LP problem exists at a vertex of the the feasible region that represents a statically determinate truss (Sheu and Schmit 1972). Hence, the optimal solution of the original NLP problem with compatibility conditions can be found by solving Problem (3.16). For a degenerate case in which the solution of Problem (3.16) is statically indeterminate, i.e., the solution of the LP problem exists on a hyperplane deﬁning the boundary of the feasible region, there exists a statically determinate truss at a vertex of the feasible region that has the same objective value as the degenerate optimal solution. Prager (1976) and Save (1983) investigated the characteristics of optimal solutions of a three-bar statically indeterminate truss. It is seen from (3.16c) that the stress constraint is obviously satisﬁed by a vanishing member with Ai = 0 and Ni = 0. This way, the discontinuity in stress constraints can be avoided for a problem with a single loading condition by simply assigning constraints on member forces instead of stresses.

Topology Optimization of Trusses

103

Note that the solution by Stolpe and Svanberg (2003b) satisﬁes all the stress constraints even for the vanishing members. We introduce the vectors of slack variables N+ = (N1+ , . . . , Ni+ )⊤ and − N = (N1− , . . . , Ni− )⊤ for N as N = N+ − N− , N+ ≥ 0, N− ≥ 0

(3.17)

For the case with σiL = −σiU , Problem (3.16) can be reformulated as an LP problem, as follows, with respect to the member forces only (Achtziger, Bendsøe, Ben-Tal, and Zowe 1992): Minimize

m ∑ 1 (Ni+ + Ni− )Li V (N , N ) = U σ i=1 i −

+

subject to D(N+ − N− ) = P Ni+

≥ 0,

Ni−

≥ 0, (i = 1, . . . , m)

(3.18a) (3.18b) (3.18c)

By minimizing V (N+ , N− ), one of the inequalities in (3.18c) is satisﬁed with equality in each member. Furthermore, the optimal truss is statically determinate, and V (N+ , N− ) turns out to be the total structural volume with Ai =

1 (N + + Ni− ) σiU i

(3.19)

The general problem including multiple loading conditions can be reformulated as a mathematical program with complementarity constraints (MPCC) (Izmailov and Solodov 2008). The stress constraints can be converted to a global constraint using a p-norm or Kreisselmeier–Steinhauser (KS) measure (Kreisselmeier and Steinhauser 1983; Wrenn 1998; Sobieszczanski-Sobieski 1992; Qiu and Li 2010). The p-norm σ P of the stress is deﬁned as P

σ =

(m ∑

) p1 σip

(3.20)

i=1

Note that σ P gives an upper bound for the maximum value of σi and converges to the maximum value as p is increased to ∞. However, an appropriately large value has to be given for p to ensure smoothness of the function. This approach has been extensively studied in the ﬁeld of continuum shape and topology optimization (Allaire, Jouve, and Maliot 2004), where the density of the material of each ﬁnite element is chosen as a design variable. The so-called solid isotropic microstructure with penalty or solid isotropic material with penalization (SIMP) approach is used (Bendsøe 1989; Rozvany, Zhou, and Birker 1992). The nonexistence of the constraint for the vanishing material density of a plate element can be incorporated by rewriting the constraint using the p-mean in conjunction with the ε-relaxed approach as (Duysinx and

104

Optimization of Finite Dimensional Structures

Bendsøe 1998; Bruggi 2008) [

})p ] p1 { m ( 1 ∑ e σi ≤1 max 0, η U + e − m i=1 ρi σ ρi

(3.21)

where ρi is the density of the ith element, m is the number of elements, σ U is the upper bound for the von Mises stress σi , and η is a penalization parameter for intermediate density. 3.5.5.2

Branch-and-bound method

For general problems with stress and displacement constraints under multiple loading conditions, reformulation of the problem to an LP problem is not possible. Therefore, a global optimization algorithm, e.g., the branch-andbound method, should be used for obtaining the optimal topology. Consider the following NLP problem under nP loading conditions: Minimize

m ∑

Ai Li

(3.22a)

i=1

subject to K(A)Uk = Pk , (k = 1, . . . , nP ) σiL ≤ σik ≤ σiU , UiL ≤ Uik ≤ UiU ,

(3.22b) P

(i = 1, . . . , m; k = 1, . . . , n )

(3.22c)

P

(3.22d)

(i = 1, . . . , n; k = 1, . . . , n )

Ai ≥ 0, (i = 1, . . . , m)

(3.22e)

where UiU and UiL are the upper and lower bounds for Ui . Ringertz (1986) proposed the following algorithm based on successive solution of the relaxed problem without stress constraints: Step 1: Assign a suﬃciently large value to the current optimal objective value V ∗. Step 2: Solve Problem (3.22) without stress constraints (3.22c) using an NLP algorithm. Stop if there is no violated stress constraint. Step 3: Solve Problem (3.22) with stress constraints for all members. Step 4: Generate a set of candidate trusses by removing each member with an active stress constraint from the optimal truss obtained in Step 3, and add them to the candidate list. Step 5: Stop if the candidate list is empty; otherwise select the last candidate in the list, and remove it from the list. Step 6: Go to Step 5 if equilibrium equations (3.22b) are not satisﬁed, i.e., if the truss is unstable.

Topology Optimization of Trusses

105

Step 7: Solve Problem (3.22) for the candidate truss without stress constraints to obtain the optimal objective value V d . If there is no violated stress constraint, then let V ∗ = min{V ∗ , V d }, and go to Step 5. If V d > V ∗ , then go to Step 5. Step 8: Solve Problem (3.22) with stress constraints to obtain the optimal objective value V s . Let V ∗ = min{V ∗ , V s }. Add new candidates to the list by removing each member with an active stress constraint from the optimal truss, and go to Step 5. Since this algorithm is rather heuristic, obtaining the global optimal solution is not guaranteed. Furthermore, because Problem (3.22) is a nonconvex problem, the solution obtained using a gradient-based NLP algorithm at Steps 3, 7, and 8 may be a local optimal solution. It is important for rapid termination of the branches of a branch-and-bound algorithm that strict valid constraints that are satisﬁed by the optimal solution are assigned to restrict the space of feasible solutions. Stolpe (2004) presented a branch-and-bound algorithm for the problem with stress, displacement, and local buckling constraints, where a formulation of simultaneous analysis and design (SAND), as shown in Sec. 1.13, is used, and the variables are the cross-sectional areas, member forces, stresses, and displacements, while the lower and upper bounds ALi and AU i , respectively, are given for Ai . Note that ALi = 0 for a topology optimization problem. He introduced valid constraints on the structural volume as well as the compliance that is formulated using a matrix inequality. The quadratic equality constraint Nik = Ai σik

(3.23)

of the ith member for the kth loading condition is relaxed to the following four linear inequalities to formulate the convex subproblems: Nik ≥ ALi σik + Ai σiL − ALi σiL , k U U U Nik ≥ AU i σi + Ai σi − Ai σi ,

Nik ≤ ALi σik + Ai σiU − ALi σiU ,

(3.24)

k L U L Nik ≥ AU i σi + Ai σi − Ai σi L For example, suppose σiL = −1, σiU = 1, ALi = 1, and AU i = 2. If Ai = Ai = 1, k k k the relation Ni = Ai σi = σi is obtained from (3.24). This strict relation is also satisﬁed for Ai = AU i = 2. However, if Ai = 1.5, the inequalities 1.5σik − 0.5 ≤ Nik ≤ 1.5σik + 0.5 are derived. Therefore, (3.24) deﬁnes an approximate relation between force and stress.

106

3.5.6 3.5.6.1

Optimization of Finite Dimensional Structures

Problem with stress and local constraints Problem formulation

In this section, the topology optimization problem under stress constraints is ﬁrst formulated as a mixed integer nonlinear programming (MINLP) problem (Kravanja, Kravanja, and Bedenik 1998). The local constraints on nodal instability and intersection of members are incorporated, and a moderately large lower bound is given for the cross-sectional area of an existing member (Ohsaki and Katoh 2005). Let yi ∈ {0, 1} denote a variable indicating by yi = 1 and yi = 0, respectively, the existence and nonexistence of the ith member of the ground structure with many nodes and members. Stress constraints should be assigned only for members with yi = 1. One of the drawbacks of the ground structure approach is that there often exist very thin members in the optimal topology, as discussed in Sec. 3.5.3, which can be prevented by assigning a moderately large lower bound ALi for the cross-sectional area of an existing member as ALi yi ≤ Ai ≤ AU (3.25) i yi , (i = 1, . . . , m) where AU i is the upper bound for Ai . Note from (3.25) that Ai = 0 should be satisﬁed if yi = 0. Another drawback in topology optimization based on the ground structure approach is that an unstable optimal truss is often obtained. A node connecting only two colinear members, as shown in Fig. 3.3, is unstable to the transverse direction of the members. An unstable solution can be avoided by introducing the lower bound for the number of members connected to an existing node. Let xr ∈ {0, 1} be the variable indicating nonexistence and existence of the rth node, respectively, by xr = 0 and xr = 1. The upper and lower bounds for the number of members connected to the rth node, if they exist, are denoted by CrU and CrL , respectively. Note that CrU is given to prevent the existence of a highly connected node. The set of indices of members connected to the rth node in the initial ground structure is denoted by Jr , and the following constraints are given: xr CrL ≤

∑

yi ≤ xr CrU , (r = 1, . . . , s)

(3.26)

i∈Jr

where s is the number of nodes including the supports. Note from (3.26) that yi = 0 should be satisﬁed by all the members connected to a removed node with xr = 0. In a practical situation, intersection of members should also be avoided, although those members are needed in the initial ground structure so as not to restrict the design space. The ith pair of mutually intersecting members in the ground structure is denoted by Si (i = 1, . . . , q). The following constraints

Topology Optimization of Trusses ∑

are to be satisﬁed:

yj ≤ 1, (i = 1, . . . , q)

107

(3.27)

j∈Si

The topology optimization problem is then formulated as an MINLP problem: Minimize

V (A) =

m ∑

Ai Li

(3.28a)

i=1 σik yi

subject to σiL yi ≤ ≤ σiU yi , (i = 1, . . . , m; k = 1, . . . , nP ) E ⊤ k d U , (i = 1, . . . , m; k = 1, . . . , nP ) σik = Li i Nik = Ai σik , (i = 1, . . . , m; k = 1, . . . , nP ) k

k

P

(3.28b) (3.28c) (3.28d)

DN = P , (k = 1, . . . , n )

(3.28e)

ALi yi ≤ Ai ≤ AU i yi , (i = 1, . . . , m) ∑ L xr Cr ≤ yi ≤ xr CrU , (r = 1, . . . , s)

(3.28f)

∑

(3.28g)

i∈Jr

yj ≤ 1, (i = 1, . . . , q)

(3.28h)

yi ∈ {0, 1}, (i = 1, . . . , m) xr ∈ {0, 1}, (r = 1, . . . , s)

(3.28i) (3.28j)

j∈Si

where the variables are A, y, x, Uk , σ k , and Nk with y = (y1 , . . . , ym )⊤ , k ⊤ ) (i = 1, . . . , m; k = 1, . . . , nP ). x = (x1 , . . . , xs )⊤ , and σ k = (σ1k , . . . , σm Note that constraint (3.28d) is nonlinear with respect to Ai and σik . If Ai takes 0 or 1, then (3.28d) can be converted to a pair of linear inequality constraints, as shown in Sec. 3.6. However, we assume that Ai has a real value. Furthermore, we cannot use the reformulation to (3.24), because the existence of singular (isolated) optimal topology is allowed here. The objective value of Problem (3.28) is denoted by V MIP . In the following, (3.28h) is called a constraint on member intersection, (3.28g) with (3.28f) is called a constraint on nodal instability, and (3.28f)–(3.28h) are referred to as local constraints. 3.5.6.2

Lower-bound and upper-bound solutions

A relaxed LP problem of Problem (3.28) is formulated by relaxing integer variables xr and yi to continuous ones and by neglecting compatibility constraints (3.28c) to ﬁnd the lower bound of V MIP . If yi = 1, (3.28b) reads σiL ≤ σik ≤ σiU , (i = 1, . . . , m; k = 1, . . . , nP )

(3.29)

and by multiplying Ai , Ai σiL ≤ Nik ≤ Ai σiU (i = 1, . . . , m; k = 1, . . . , nP )

(3.30)

108

Optimization of Finite Dimensional Structures

is derived. Note that (3.30) is satisﬁed also for yi = 0 with Ni = 0 because constraint (3.28f) is imposed in the relaxed problem. Therefore, (3.30) is satisﬁed for 0 ≤ yi ≤ 1 if (3.28b) is satisﬁed. Hence, the relaxed LP problem of MIP (3.28) is formulated as Minimize

V (A) =

m ∑

Ai Li

(3.31a)

i=1 k

subject to DNk = P , (k = 1, . . . , nP ) Ai σiL ALi yi

≤ ≤ Ai σiU , ≤ Ai ≤ AU i yi ,

xr CrL ≤ ∑

Nik

∑

(3.31b) P

(i = 1, . . . , m; k = 1, . . . , n ) (i = 1, . . . , m)

yi ≤ xr CrU , (r = 1, . . . , s)

(3.31c) (3.31d) (3.31e)

i∈Jr

yj ≤ 1, (i = 1, . . . , q)

(3.31f)

0 ≤ yi ≤ 1, (i = 1, . . . , m) 0 ≤ xr ≤ 1, (r = 1, . . . , s)

(3.31g) (3.31h)

j∈Si

where the variables are A, y, x, and Nk . Note again that Problem (3.31) is an LP problem, where the global optimality of the solution is guaranteed. If a statically determinate truss satisfying the local constraints is obtained by solving LP (3.31), then the solution gives the global optimal topology of the original MIP problem. If a statically indeterminate truss is obtained, then the objective value V LP of LP (3.31) is a lower bound of the true optimal objective value of MIP (3.28), because the solution of MIP (3.28) satisﬁes all the constraints of LP (3.31). For a given set I of the indices of existing members of the optimal solution of Problem (3.31), i.e., I = {j | yj > 0}, the following NLP, called NLPI , is deﬁned. If NLPI is feasible, then its objective value gives an upper bound V NLPI of V MIP , because the solution of NLPI satisﬁes all the constraints of MIP (3.28): ∑ Minimize V (A) = Ai Li (3.32a) subject to

i∈I L k σi ≤ σi (A) ≤ σiU , ALi ≤ Ai ≤ AU i , (i

(i ∈ I; k = 1, . . . , nP ) ∈ I)

(3.32b) (3.32c)

where the variables are A only. 3.5.6.3

Branch-and-bound method

Using the upper- and lower-bound solutions, the original MINLP problem (3.28) is solved using the standard branch-and-bound method, which is summarized as follows:

Topology Optimization of Trusses

109

Step 0: Initialize the upper bound V U as V U = ∞. Let the set A of the active problems consist of the original MINLP problem (3.28). Step 1: Select a problem P from A and remove it from A. Step 2: Solve the relaxed LP problem P of P , and select a member i from I with 0 < yi < 1 of the solution. Let P0 and P1 denote the subproblems of P by specifying yi = 0 and 1, respectively. Solve the relaxed LPs (3.31) of P0 and P1 , which are denoted by P 0 and P 1 , respectively. Step 3: Let V0 and V1 denote the optimal objective values of P 0 and P 1 , respectively. If V0 > V U , set yi = 1 and terminate P0 . If V1 > V U , set yi = 0 and terminate P1 . Step 4: If the topology conforming to the LP solution satisﬁes all the local constraints, then compute the objective value of NLPI (3.32) denoted by V NLPI after specifying the set I of the existing members. Step 5: If V NLPI < V U , then update V U to V NLPI . Step 6: If V0 and/or V1 is less than V U , then add P0 and/or P1 , respectively, to A and go to Step 1. Step 7: If A is not empty, then go to Step 1. Step 8: Output the best value of V U . Compute the lower bound V L backward from the bottom of the branching tree so that V L of the parent problem P is updated by min{V0 , V1 } if V L < min{V0 , V1 }. The basic branch-and-bound algorithm is presented in Appendix A.2.5. At the ﬁrst stage of the branching process, we can obtain an upper bound V U by ﬁxing yi appropriately for all the members existing in the lower-bound solution so that the local constraints are satisﬁed. Furthermore, we can use the local constraints before solving the LP problem (3.31); i.e., xr can be ﬁxed at 1 if a member exists such that yi = 1 (i ∈ Jr ). Alternatively, xr is equal to 0 if yi = 0 for all the members in Jr . Furthermore, yi (i ∈ Sr ) can be ﬁxed at 0 if there exists a member with yj = 1 (j ∈ Sr , j ̸= i). Because Problem (3.32) is nonconvex, the global optimum of the solution obtained by this branch-and-bound algorithm cannot be guaranteed. However, in the following numerical examples, it is conﬁrmed that the same optimal solutions can be found starting from diﬀerent initial solutions. Therefore, the nonconvexity of Problem (3.32) is not very strong, and the global optimal solution can be found for almost all cases. 3.5.6.4

Numerical examples

Optimal topologies of plane trusses are found using the branch-and-bound algorithm. In the following examples, the units of force and length are omitted

110

Optimization of Finite Dimensional Structures (1)

1

18

24

3

12

(7) 26 5

20 (5)

(2) 2

(4)

9

10

(8) 23

29

y 11

x

(6)

14

(9)

P

17

7 (11)

16 27 6

21 4

(3)

(10) 28

22

13 25

19

15

8 (12)

P

FIGURE 3.9: A 3 × 2 plane grid (from Ohsaki and Katoh, Struct. Mulc tidisc. Optimiz., 29, 190–197, 2005, Fig. 2, Copyright ⃝Springer Science+Business Media, reprinted with kind permission).

for brevity. Problem (3.31) is solved by HOPDM Ver. 2.13 (Gondzio 1995), which utilizes the higher-order primal-dual interior-point method. The solutions of NLP problems are found using NLPQL implemented as DNCONG in the IMSL library (Visual Numerics Inc. 1997), where sequential quadratic programming is used. Sensitivity coeﬃcients of stresses and displacements with respect to the cross-sectional areas are computed using the direct diﬀerentiation method; see Sec. 2.2 for details. As a simple example, consider the ﬁve-bar truss in Fig. 3.6. The bounds of the cross-sectional areas are ALi = 0 and AU i = 20.0. The local constraints are not given for comparison purposes with the results of Cheng and Guo (1997). However, from stability requirements, CrL should be 1 for the supports and 2 for the loaded nodes, and xr should be 1 for all the nodes and supports. Other parameters are the same as in Example 3.2 in Sec. 3.5.3. The initial LP solution has intersecting members 3 and 4, and member 4 is selected as the branching member. The optimal objective value of P 0 is 33.500. Since the optimal truss of P 0 is statically determinate, V U is updated to 33.500. The optimal objective value of P 1 is 32.500. All the members exist in the solution of P 1 , and the value of V NLPI is 39.986. By solving LP problem (3.31) and NLP problem (3.32) 15 times and 4 times, respectively, we obtain the best upper bound solution (A1 , A2 , A3 , A4 , A5 ) = (1.0000, 2.5000, 10.0000, 0.0, 14.1421) with V U = 33.500, which agrees with the result by Cheng and Guo (1997). The lower bound V L is 32.500, which is slightly smaller than V U . Next we consider plane grid trusses with square units. A truss with a 3 × 2 grid is shown in Fig. 3.9. The lengths of the members in the x- and y-

Topology Optimization of Trusses

111

4 5

3 1 2

(a) Initial LP solution.

(c) Final upper-bound solution.

(b) Initial upper-bound solution.

(d) Final lower-bound solution.

FIGURE 3.10: Optimal solutions of 2 × 2 plane grid (from Ohsaki and Katoh, Struct. Multidisc. Optimiz., 29, 190–197, 2005, Figs. 3–6, Copyright c ⃝Springer Science+Business Media, reprinted with kind permission).

directions are 200. Irrespective of the number of grids, two loading conditions are considered, where the loads in the negative y-directions are applied at the node at the lowest end (node 12 in Fig. 3.9) and the node left of the lowest end (node 9 in Fig. 3.9), respectively. The magnitude of each load is 1000. The lower- and upper-bound cross-sectional areas are given as ALi = 200 and U AU i = 800. The bounds for the stress are ±2.0, and Cr = 6. The value L of Cr is 1 for the supports, 2 for the node at the lowest end, and 3 for the remaining nodes. Note that these lower bounds are deﬁned naturally from the requirements of stability and equilibrium, and they do not unnecessarily restrict the design space. Optimal topology is ﬁrst found for the 2×2 grid. The LP solution at the ﬁrst step is as shown in Fig. 3.10(a), where the width of a member is proportional to its cross-sectional area. The objective value V LP is 7.0000 × 106 . To obtain an initial upper-bound solution, member 1 in Fig. 3.10(a) is removed because it has the smaller cross-sectional area in the pair of intersecting members. Note that this selection is heuristic; however, only a good upper bound is to

112

Optimization of Finite Dimensional Structures

TABLE 3.1:

Optimization results of plane grid trusses (from Ohsaki and Katoh, Struct. Multidisc. Optimiz., 29, 190–197, 2005, Table 1, Copyright c ⃝Springer Science+Business Media, reprinted with kind permission). No. of division 2×2 3×2 3×3 4×4 4×4 4×4

(m, n)

(ALi , AU i )

(20,14) (29,20) (42,28) (72,46) (72,46) (72,46)

(200,800) (200,800) (200,800) (200,800) (200,600) (400,800)

No. of steps 121 942 5874 64890 68656 41001

No. of LP 64 571 3483 42831 42707 26580

No. of NLP 5 6 23 7 73 3

VU ×107 0.7900 1.2900 1.2467 1.7067 1.8373 2.1507

VL ×107 0.7800 1.2800 1.2467 1.7067 1.7916 2.1507

be found at this stage. After removing member 1, the node connected by members 2 and 3 is removed because it violates the local constraint (3.28g) with CrL = 3, and, accordingly, members 4 and 5 are to be removed. NLP problem (3.32) is solved by ﬁxing the values of yi and xr to ﬁnd an upper bound solution, as shown in Fig. 3.10(b), where V NLPI is 8.0000 × 106 . The branch-and-bound process is carried out to ﬁnd the ﬁnal upper-bound solution, as shown in Fig. 3.10(c), where V U = 7.9000 × 106 . The optimization results are listed in the ﬁrst row of Table 3.1, where No. of steps means the number of diﬀerent topologies that have been searched. In this example, only 121 topologies are searched out of the total of 220 ≃ 106 possible topologies. The ﬁnal lower-bound solution is as shown in Fig. 3.10(d), where V L = 7.8000 × 106 . Since this truss is statically indeterminate, the axial forces obtained by solving LP problem (3.31) are not correct. The maximum absolute value of the ratio of stress to the upper or lower bounds is 1.1111 if the compatibility conditions are considered; i.e., the solution does not satisfy stress constraints. Hence, V L has a smaller value than V U . However, the diﬀerence between V L and V U is very small and the good upper-bound solution has been found after solving NLP problem (3.32) only ﬁve times. If we do not incorporate the local constraints, the numbers of steps, LP, and NLP are 529, 350, and 58, respectively. The optimization results for 3 × 2, 3 × 3, and 4 × 4 grids are also listed in Table 3.1. The ﬁnal upper-bound solution for a 4×4 grid is shown in Fig. 3.11. Note that V L = V U is satisﬁed for 3 × 3 and 4 × 4 grids, because the lowerbound solutions are statically determinate. The number of LP steps increases drastically as the size of the problem, such as the number of members, is increased. The number of NLP steps, however, is independent of the problem size, because it depends on the quality of the initial upper-bound solution. If AU i is decreased to 600 for the 4 × 4 grid, the number of NLP steps increases to 73, as shown in the ﬁfth row of Table 3.1. If we do not consider the local constraints for the 3 × 2 truss, the number of steps, LP, and NLP are

Topology Optimization of Trusses

113

FIGURE 3.11: Final upper-bound solution for the 4 × 4 plane grid (from Ohsaki and Katoh, Struct. Multidisc. Optimiz., 29, 190–197, 2005, Fig. 9, c Copyright ⃝Springer Science+Business Media, reprinted with kind permission).

2364, 1468, and 391, respectively. We can observe from these results that the computational cost can be drastically reduced by using local constraints.

3.6 3.6.1

Mixed integer programming for topology optimization with discrete variables Introduction

As we have seen in Secs. 3.3 and 3.5, the most popular approach to topology optimization of trusses is the ground structure approach, where the crosssectional areas are regarded as continuous variables. However, in practical situations of civil and architectural engineering, the cross-sectional properties are selected from the discrete list of available standard sections. Furthermore, standard layouts are often used for plane and spatial trusses. Therefore, it is very important to investigate the optimal topology as an assembly of standard layouts, e.g., Schwedler dome and Lamella dome for spatial trusses, and Howe truss and K-truss for bridge-type plane trusses (Narayanan 2006). In Sec. 3.5, we presented a mixed integer nonlinear programming (MINLP) approach to incorporate local constraints with continuous cross-sectional areas. In this section, we do not consider local constraints; however, we formulate the problem as an MINLP problem by considering the cross-sectional areas as discrete variables. An example of relaxation of the compliance minimization problem to a convex quadratic programming problem by Achtziger and Stolpe (2008, 2009) is ﬁrst presented. Next, the problem under stress con-

114

Optimization of Finite Dimensional Structures

straints is reformulated to an MIP problem following the approach of Stolpe and Svanberg (2003a), and we present the results of truss topology optimization as an assembly of standard layouts (Watada and Ohsaki 2009c).

3.6.2

Compliance minimization problem

Consider ﬁrst a problem of minimizing the compliance under constraint on total structural volume. Let Ai and Li denote the cross-sectional area and length of the ith member, respectively. Ai is selected from the set of available sections as Ai ∈ {0, a1i , . . . , ari i } (i = 1, . . . , m), where m is the number of members, and ri is the number of available sections for the ith member if it exists. The ith member does not exist if Ai = 0. The truss is subjected to the nodal loads P. The displacement vector and the stiﬀness matrix are denoted by U and K(A), respectively. Using the approach of simultaneous analysis and design (see Sec. 1.13), the optimization problem is formulated as

Minimize P⊤ U subject to K(A)U = P m ∑ Ai Li ≤ V U

(3.33a) (3.33b) (3.33c)

i=1

Ai ∈ {0, a1i , . . . , ari i }, (i = 1, . . . , m)

(3.33d)

where the variables are A and U, and V U is the speciﬁed upper bound of the total structural volume. Problem (3.33) is a nonconvex quadratic programming (QP) problem with mixed continuous and discrete variables, which can be solved using a branchand-bound method if the problem relaxing Ai to a continuous variable can be formulated as a convex programming problem. For this purpose, Achtziger and Stolpe (2008, 2009) presented two types of relaxed problems. Let λ denote an auxiliary variable. The lower and upper bounds for the continuous L k variable Ai are denoted simply by ALi and AU i , respectively; e.g., Ai = ai and k+1 U 0 Ai = ai , 1 ≤ k ≤ ri − 1 with ai = 0 in the process of branch-and-bound. Using the obvious relation

m ∑ i=1

ALi ≤

m ∑ i=1

Ai ≤

m ∑ i=1

AU i

(3.34)

Topology Optimization of Trusses

115

a relaxed problem can be formulated as

Minimize

⊤

− P U + λV

U

µi

(3.35a)

i=1

) 1 U⊤ Ki U − λ ALi + µi ≤ 0, (i = 1, . . . , m) 2Li ( ) 1 U⊤ Ki U − λ AU i + µi ≤ 0, (i = 1, . . . , m) 2Li (

subject to

−

m ∑

λ≥0

(3.35b) (3.35c) (3.35d)

where the variables are µ1 , . . . , µm , U, and λ, and Ki is deﬁned by

K=

m ∑

Ai Ki

(3.36)

i=1

Here, notations consistent with those in this book have been used, while the volume of each member is taken as a design variable by Achtziger and Stolpe (2008). Note that Problem (3.35) is a convex quadratic programming problem, for which the global optimal solution can be found using any method of nonlinear programming. However, the number of variables is very large if members exist between all pairs of nodes in the ground structure with many nodes. Therefore, Achtziger and Stolpe (2008, 2009) presented another mildly convex relaxed problem that can be solved eﬃciently with small computational cost.

3.6.3

Stress constraints

Suppose again the cross-sectional area Ai of the ith member is selected from the set of available sections Ai ∈ {0, a1i , . . . , ari i }. Then the axial force vector N = (N1 , . . . , Nm )⊤ is given with respect to U using the constant vector bi , elastic modulus E, and the relations in (3.8) as

Ni =

Ai E ⊤ b U, (i = 1, . . . , m) Li i

(3.37)

Let σiU and σiL denote the upper and lower bounds for the stress of the ith member. The equilibrium matrix is denoted by D, of which the ith column is equal to bi . Then the topology optimization problem for minimizing the

116

Optimization of Finite Dimensional Structures

total structural volume under stress constraints is formulated as Minimize

V =

m ∑

Ai Li

(3.38a)

i=1

subject to DN = P

(3.38b)

≤ Ni ≤ (i = 1, . . . , m) E ⊤ Ni = Ai bi U, (i = 1, . . . , m) Li Ai ∈ {0, a1i , . . . , ari i }, (i = 1, . . . , m) Ai σiL

Ai σiU ,

A 0–1 variable xki is deﬁned as { 1 if Ai = aik k xi = , (i = 1, . . . , m; k = 1, . . . , ri ) 0 otherwise

(3.38c) (3.38d) (3.38e)

(3.39)

Then Ai is written with aki and xki as ri ∑

Ai =

xki aki

(3.40)

k=1

with

ri ∑

xki ≤ 1

(3.41)

k=1

Note that (3.40) and (3.41) indicate that at most only one of xki (k = 1, . . . , ri ) can take 1; if all of them are equal to 0, then Ai = 0 and member i is removed. Hence, a variable xi , indicating existence and nonexistence of member i by xi = 1 and 0, respectively, is deﬁned as xi =

ri ∑

xki

(3.42)

k=1

An auxiliary variable ski is also used for the axial force as ski = xki aki Ni =

ri ∑

E ⊤ b U Li i

ski

(3.43a) (3.43b)

i=1

Note that (3.43a) is nonlinear with respect to the variables xki and U. Let Umin = (U1min , . . . , Unmin )⊤ and Umax = (U1max , . . . , Unmax )⊤ denote the vectors of lower and upper bounds of displacements, where n is the number of degrees of freedom. Suﬃciently small lower bound cmin and large upper i bound cmax are given for the elongation of member i corresponding to the i

Topology Optimization of Trusses

117

displacements satisfying Umin ≤ U ≤ Umax . Then the nonlinear relation (3.43a) is converted to the following pair of inequalities (Stolpe and Svanberg 2003a; Glover 1975): (1 − xki )cmin ≤ aki i

E ⊤ b U − ski ≤ (1 − xki )cmax i Li i

(3.44)

It is seen from (3.44) that if xki = 1, then the left-hand side and right-hand side terms vanish, and, accordingly, relation (3.43a) is satisﬁed. In contrast, if xki = 0, then the inequalities hold for any value of U satisfying Umin ≤ U ≤ Umax ; i.e., no constraint is given between U and ski . Additional requirements are to be assigned to obtain practical optimal solutions, for example (Ohsaki and Watada 2008; Watada and Ohsaki 2009b, 2009c): • Cond-1: Intersecting members i and j in the ground structure cannot exist simultaneously in the optimal topology: xi + xj ≤ 1

(3.45)

• Cond-2: If member i exists, then member j exists: xj ≥ xi

(3.46)

• Cond-3: If member i does not exist, then member j does not exist: xj ≤ xi

(3.47)

• Cond-4: Cond-2 and Cond-3 are simultaneously satisﬁed: xj = xi

(3.48)

In the numerical examples, the members are classiﬁed into groups to represent the standard conﬁgurations. For this purpose, auxiliary variables yk ∈ {0, 1} (k = 1, . . . , ny ) are used for indicating by yk = 1 and 0 the existence and nonexistence, respectively, of the kth group, where ny is the number of groups. Constraints for yk are formulated as follows: • Cond-5: Members i and j are included in group k; i.e., these members exist (do not exist) if group k exists (does not exist): yk = xj = xj

(3.49)

• Cond-6: If group k exists, then group j exists: yj ≥ yk

(3.50)

118

Optimization of Finite Dimensional Structures

• Cond-7: If group k does not exist, then group j does not exist: yj ≤ yk

(3.51)

• Cond-8: If Cond-6 and Cond-7 are simultaneously satisﬁed, we make another group l containing groups j and k: yl = yj = yk

(3.52)

These constraints are symbolically written as follows using the vectors x = (x1 , . . . , xm )⊤ , y = (y1 , . . . , yny )⊤ , and matrices Gx , Gy , Hx , and Hy : Gx x + Gy y = 0,

(3.53)

Hx x + Hy y ≤ 0

The auxiliary variables and constraints can be systematically generated using a decision tree (Watada and Ohsaki 2009c). Finally, the topology optimization problem for minimizing the total structural volume V is formulated as Minimize

V =

ri m ∑ ∑

xki aki Li

(3.54a)

i=1 k=1

subject to DN = P ri ∑ Ni = ski , (i = 1, . . . , m) i=1 xki aki σiL ≤ ski ≤ xki aki σiU , (i = 1, . . . , m; Ujmin ≤ Uj ≤ Ujmax , (j = 1, . . . , n)

(1 − xki )cmin ≤ aki i ri ∑

(3.54b) (3.54c) k = 1, . . . , ri ) (3.54d)

E ⊤ b U − ski ≤ (1 − xki )cmax , i Li i (i = 1, . . . , m; k = 1, . . . , ri )

xki ≤ 1, (i = 1, . . . , m)

k=1 x

(3.54e)

(3.54f) (3.54g)

G x + Gy y = 0 Hx x + Hy y ≤ 0

(3.54h) (3.54i)

xki ∈ {0, 1}, (i = 1, . . . , m; k = 1, . . . , ri ) yi ∈ {0, 1}, (i = 1, . . . , ny )

(3.54j) (3.54k)

which is an MIP problem with variables xki , ski (i = 1, . . . , m; k = 1, . . . , ri ), Uj (j = 1, . . . , n), x, y, and N.

Topology Optimization of Trusses

119

P

FIGURE 3.12:

A bridge-type plane truss.

(a) Howe truss

(b) Platt truss

(c) Warren truss

FIGURE 3.13:

3.6.4

(d) K-truss

Four standard layouts of plane trusses.

Numerical examples

Consider a bridge-type plane truss, as shown in Fig. 3.12, which represents the ground structure that has many members as an assemblage of the four standard layouts, namely, Howe truss, Platt truss, Warren truss, and K-truss, as shown in Figs. 3.13(a)–(d), respectively. Uniform vertical loads P are applied at the lower nodes. Incorporating the symmetry conditions of the loads and topology of the ground structure, the optimal topology of the truss is also assumed to be symmetric; hence, one of the half parts is to be optimized. We use the optimization library CPLEX 10.2 (ILOG 2007) for solving the MIP. Computation is carried out on a PC with Intel Xeon CPU 3.40GHz, 2.00GB RAM. The units of force and length are omitted for brevity. The members are appropriately divided into groups so that one of the four types is selected for each of the 12 units of the half part. For this purpose, the members in a unit are classiﬁed into three groups, as shown in Fig. 3.14. The parameters are H = 2.0, W = 12.0, E = 2.0, and σiU = 0.003 for all members. The lower-bound stress is given as follows to incorporate the eﬀect of member buckling: σiL = −

L2min U σ L2i i

(3.55)

where Lmin is the length of the shortest member in the ground structure. The cross-sectional areas are chosen from the list {0.0, 1.0, 1.8}; i.e., a1i = 1.0, a2i = 1.8, and ri = 2 for all the members. In order to improve computational

120

FIGURE 3.14:

Optimization of Finite Dimensional Structures

Classiﬁcation of members in a unit.

(a) Case 1: CPU = 3500, V = 65.410

(b) Case 2: CPU = 960, V = 80.810

(c) Case 3: CPU = 380, V = 82.414

FIGURE 3.15: truss.

Optimal topologies and cross-sectional areas of the plane

eﬃciency, the variable ski is scaled to sk∗ i as sk∗ i =

Li k s Eakj i

(3.56)

The optimal solutions are found for Cases 1–3 with three diﬀerent load magnitudes P = 0.00007, 0.00012, and 0.00014, respectively. The optimal topologies are as shown in Figs. 3.15(a)–(c), where the width of each member is proportional to its cross-sectional area, and the CPU time (s) and the optimal objective value are shown in each ﬁgure. As is seen, the Warren truss tends to be selected for a smaller load, while the Howe truss is selected for a larger load. Because vertical members are allowed to be removed in the Warren truss, it has the smallest member density among the four types, and, accordingly, is selected for the case of a smaller load. Note that the CPU time strongly depends on the load magnitude; i.e., computational cost is smaller for a larger load, because the number of admissible sets of cross-sectional areas decreases as the load is increased. We next optimize a dome truss as an assemblage of the Schwedler dome and Lamella dome, as shown in Figs. 3.16(a) and (b), respectively, subjected to the vertical load P at the top node. The cross-sectional areas are selected from the list {0.0, 1.0, 3.0, 5.0}, and we consider Cases 1–3 with loads P = 0.001, 0.002, and 0.0025, respectively. The whole structure is solved without considering the symmetry condition. The optimal topologies are as shown in Figs. 3.17(a)– (c). As is seen, the optimal truss is a Lamella dome for Case 1 with the

Topology Optimization of Trusses

(a) Schwedler dome

FIGURE 3.16:

121

(b) Lamella dome

Two standard layouts of dome trusses.

(a) Case 1: CPU = 12, V = 1404.3

(b) Case 2: CPU = 688, V = 1441.1

(c) Case 3: CPU = 680, V = 1474.4

FIGURE 3.17:

Optimal topologies of the dome truss.

smallest load, while the Schwedler dome dominates for larger load magnitudes. Note that the cross-sectional areas are 1.0 for all the existing members except those connected to the center nodes for Cases 2 and 3, which are equal to 3.0 and 5.0, respectively. In this example, there is no correlation between the load magnitude and the CPU time.

122

Optimization of Finite Dimensional Structures

3.7

Genetic algorithm for truss topology optimization

3.7.1

Introduction

The genetic algorithm (GA) has been successfully applied to optimization problems in various ﬁelds, including computer science, social science, and operations research. Although the GA was developed by Holland (1975), its concept have said to been proposed by Rechenberg (1965). GAs are very useful for problems with discontinuous cost and/or constraint functions because the gradients of the functions are not needed for ﬁnding approximate optimal solutions (Goldberg 1989). Since the 1990s, GAs have been widely applied to structural optimization problems (Jenkins 1991; Hajela and Lin 1992; Rajeev and Krishnamoorthy 1997; Hayalioglu 2000; Ant´ onio 2002). Grierson and Pak (1993) applied a GA to a frame topology optimization problem, where the lists of topologies and cross-sectional properties as well as the nodal locations are encoded into binary strings. The null cross-sectional area is also included in the list to enable the removal of members. Hajela and Lee (1995) presented a two-level approach to ensure the stability of the truss. They also presented a scaling approach for constrained optimization problems, where the increase of the penalty term is limited to avoid too much penalty compared with the objective value. GAs are extensively used for optimization of composite laminas. For instance, as an early development, Hajela and Lin (1991) optimized the stacking sequence of a composite beam with dampers. Marcelin, Trompette, and Dornberger (1995) developed a multiobjective GA for design of composite beams. Kogiso, Watson, G¨ urdal, and Haftka (1994) presented a binary tree approach for checking duplicate solutions as well as an operator for local improvement of composite laminates. Kameshki and Saka (2003) optimized a frame with a semi-rigid connection using GA. Kocer and Arora (1999) compared the performances of GA and simulated annealing for optimization of transmission towers. GA is very useful for design under constraints on inelastic responses, for which the sensitivity coeﬃcients of responses are diﬃcult to obtain (Yun and Kim 2005). There are several variants of GAs, such as the evolutionary algorithm (EA). Greiner, Emperador, and Winter (2004) applied an improved EA with a rebirth operator to single and multiobjective optimization of frames. Giger and Ermanni (2006) presented a graph-based approach, where the operators of mutation and crossover are introduced for the adjacency matrix deﬁning the topology of the truss. In this section, an example of a GA is presented for topology optimization of trusses with stress constraints, considering the nodal costs (Ohsaki 1995).

Topology Optimization of Trusses

3.7.2 3.7.2.1

123

Optimization considering nodal cost Problem formulation

In the practical design process of frames and trusses, the cost of the nodes connecting the members may be sometimes equivalent to or greater than that of the members. Therefore, the nodal cost should be included in the cost (objective) function to obtain more realistic optimal topology. A standard ground structure approach is used here. A node should exist if at least one member connected to the node has a positive cross-sectional area, and the size of the node is selected from a list of predeﬁned values. Therefore, the nodal cost is a discontinuous function of the cross-sectional areas, which are considered as design variables, and no algorithm utilizing design sensitivity analysis can be applied even if the cross-sectional areas are continuous variables. The length and the cost per unit volume of the ith member are denoted by Li and ci , respectively. Let ai (A) denote the maximum value of the crosssectional areas among the members connected to the ith node. It is assumed here, for a simple presentation of the algorithm, that only one type of node is used; i.e., the cost gi (A) of the ith node is deﬁned as { gi (A) = g for ai (A) > 0 (3.57) gi (A) = 0 for ai (A) = 0 where g is the prescribed cost of a node. Let n and nS denote the number of degrees of freedom and the number of nodes, respectively. The truss is subjected to nP sets of static loads, and the value corresponding to the kth load set is denoted by the superscript ( · )k . The upper and lower bounds for the jth displacement component Ujk are denoted by UjU and UjL , respectively. The bounds for the stress σik of the ith member are deﬁned similarly. The topology optimization problem for minimizing the total cost C(A) is formulated as Minimize subject to

C(A) =

m ∑

n ∑ S

ci Ai Li + gi (A) i=1 i=1 UjL ≤ Ujk ≤ UjU , (j = 1, . . . , n; σiL ≤ σik ≤ σiU for Ai > 0,

(3.58a) k = 1, . . . , nP )

(i = 1, . . . , m; k = 1, . . . , nP )

(3.58b) (3.58c)

Note from (3.58c) that the stress constraints are assigned only for the existing members.

3.7.3

Topological bit and ﬁtness function

In order to solve Problem (3.58) using a GA, the cross-sectional area of each member is represented by a string of le binary data. Let A0 denote the unit

124

Optimization of Finite Dimensional Structures

value of the cross-sectional area. For the case le = 4, a string 1010 for the ith member, for instance, indicates Ai = (1 × 23 + 0 × 22 + 1 × 21 + 0 × 20 ) × A0 = 10A0 . A topological bit Gei is added to the left of the string representing Ai to indicate the existence/nonexistence of member i. A member is to be removed if Ai = 0 or Gei = 0. For example, members with string 01001 (Ai = 9A0 , Gei = 0) and 10000 (Ai = 0, Gei = 1) are to be removed. Although the GA is a heuristic approach without any strong theoretical basis, the schema theorem is usually used for explaining the convergence to an optimal solution by the GA. A pattern that has a certain meaning in a one-dimensional bit string is called a schema or building block. For example, a schema H = 01 ∗ ∗1 ∗ ∗ means that the ﬁrst bit is ﬁxed at 0, the second and ﬁfth bits should be equal to 1, and the other bits can be either 0 or 1 as indicated by the wildcard ∗. Let δ(H) denote the deﬁning length of the schema H that is the distance between the ﬁrst and the last bits with ﬁxed values. The number of bits with ﬁxed values in H is called the order, which is denoted as o(H); e.g., δ(H) = 4 and o(H) = 3 for the schema H = 01 ∗ ∗1 ∗ ∗. Let M (H, t) denote the number of individuals that contain the schema H at the generation t. The mean ﬁtness value of the individuals containing H is denoted by f (H), whereas the mean ﬁtness value of all individuals is denoted by f . If we do not consider mutation or crossover, then the expected number of individuals containing H at the generation t + 1 is given as M (H, t + 1) = M (H, t)

f (H) f

(3.59)

Let l denote the total length of the string. The probabilities of crossover and mutation are denoted by pc and pm , respectively. Then the probabilities of the schema H to be destroyed by mutation and crossover are given as o(H)pm and pc δ(H)/(l − 1), respectively. Because H may be destroyed by crossover and mutation simultaneously, the expected number of individuals containing H at the generation t + 1 is given with inequality as ( ) f (H) δ(H) M (H, t + 1) ≥ M (H, t) 1 − pc − o(H)pm (3.60) l−1 f It is seen from (3.60) that the schema with the larger ﬁtness value f (H), i.e., f (H)/f > 1, will dominate exponentially with the progress of generation. For example, for our problem, consider a case where the nonexistence of member 1 leads to a larger ﬁtness value. If the topological bit Gei is not incorporated, the solutions without member 1 are represented by a schema Ha = 0000 ∗ ∗ ∗ . . . , which leads to δ(Ha ) = 3 and o(Ha ) = 4. If Gei is incorporated at the left of the string representing the cross-sectional area, the schema corresponding to Ge1 = 0 is Hb = 0 ∗ ∗ ∗ . . . , i.e., δ(Hb ) = 0 and o(Hb ) = 1; hence, the probability of destroying the building block that leads to a larger ﬁtness value is reduced. Therefore, from the schema theorem, incorporation of the topological bit leads to rapid convergence to optimal

Topology Optimization of Trusses

125

topology. If the nodal cost is moderately large, a truss with small number of members has a large ﬁtness value, and such optimal topology can easily be found by incorporating the topological bit. In order to prevent numerical diﬃculties for computing the displacements and stresses of the truss, a small value ALi is assigned for the cross-sectional area of a member with Ai = 0 or Gei = 0. The member with Ai = ALi is to be removed from the ﬁnal optimal solution. Hence, aj (A) in the conditions (3.57) is redeﬁned as the maximum value of Ai − ALi among the members connected to node j in the ground structure. The constraint (3.58c) is also modiﬁed as σiL ≤ σik ≤ σiU for Ai > ALi , (i = 1, . . . , m; k = 1, . . . , nS )

(3.61)

A random number 0 ≤ τ < 0 is generated to assign the value of each topological bit at the initial generation. If τ is less than the prescribed probability pe of existence, then Gei = 1; otherwise Gei = 0. The ﬁtness function is deﬁned as ) ( n m nS nP m ∑ ∑ ∑ ∑ ∑ ∗ D S C = C0 − ci Ai Li − gi − dkj + dki (3.62) i=1

i=1

k=1

i=1

i=1

where C0 is a large positive value to prevent the existence of an individual with a negative ﬁtness value. Note that the value of C0 does not have any eﬀect on the optimal solution if a ranking strategy is used. The last term in (3.62) is the penalty for violating the stress and displacement constraints, S D S U L U L where dD kj and dki are deﬁned with parameters C , C , eD , eD , eS , and eS as (

dD kj dSkj

)2 ( )2 Ujk Ujk L 1− U =C + + eD 1 − L Uj Uj ( ) ( )2 2 σik σik S U L + eS 1 − L = C + eS 1 − U σi σi D

eU D

(3.63a) (3.63b)

L U L Note that eU D , eD , eS , and eS have positive values, respectively, if the corresponding constraints on the responses are violated; otherwise they are 0. L D Furthermore, C D > 0 only if eU = 0. C S is D > 0 or eD > 0; otherwise C deﬁned similarly. The ranking strategy is used here to prevent the dominance of overly ﬁt designs in the early generation, and to ensure moderate convergence in the ﬁnal stage. Let Rj denote the rank of the ith individual (solution), and deﬁne the ﬁtness function Vj as Vj = (Rj − nS )2 (3.64)

where Rj = 1 for the individual with the smallest value of C ∗ . Then the following simple selection strategy is used to calculate the probability pj of

126

Optimization of Finite Dimensional Structures 2m

2m

2m Px Py

FIGURE 3.18:

A 20-bar plane truss.

(a) Case 1: C = 311.67 FIGURE 3.19:

2m

(b) Case 2: C = 280.46

Optimal topologies of the 20-bar plane truss.

reproduction: ∑ Vj Vk , V = V k=1 s

pj =

(3.65)

where s is the number of individuals in a generation (size of the population).

3.7.4

Numerical examples

Optimal topologies are found for a 20-bar plane truss, as shown in Fig. 3.18, under stress constraints to discuss the eﬀectiveness of the use of the topological bit. The truss is subjected to the two loading conditions (Px , Py ) = (0, 10000 N) and (7071.07 N, −7071.07 N). The parameters are A0 = 50.0 mm2 , s = 150, ci = 5.0 × 10−6 mm−3 for all members, g = 25.0, pc = 1, and pm = 0.01. It is assumed here that the cost of a support is the same as that of 2 a node. The bounds for stress are σiU = −σiL = 98.0 N/mm for all members. L A very small value Ai is given for the cross-sectional area of a removed

Topology Optimization of Trusses

127

600 Case 1 Case 2

550

Total cost

500 450 400 350 300 250 0

5

10

15

20

25

30

35

40

45

50

Generation

FIGURE 3.20: History of total cost C of the most ﬁt design of the 20-bar plane truss.

member with Ai = 0 or Gei = 0. Therefore, there may exist nearly unstable design Ai = ALi for all members in the optimization process if only the stress constraints are considered. Note that such a design has a large ﬁtness value, because the member cost is extremely small and the stress constraint is allowed to be violated in all members. Furthermore, the nodal cost for the design Ai = ALi for all members is 0. Because the displacements of this solution are very large, the convergence to this trivial optimal solution can be successfully prevented by imposing constraints on displacements. Hence, an artiﬁcial upper bound UjU = 10.0 mm is given for each displacement component. Optimal topologies are found for two cases with pe = 1.0 and 0.5, which are denoted as Cases 1 and 2, respectively. Note that pe = 1.0 in Case 1 means that the topological bit is not used. Fig. 3.19(a) shows the optimal topology for Case 1. The truss has six nodes and nine members, and the total cost C is 311.67 with member cost 161.67. The optimal topology for Case 2 is as shown in Fig. 3.19(b). Five nodes and six members exist in this truss, and the total cost C and the total member cost are 280.46 and 155.46, respectively. As is seen, optimal topologies with a small number of members can be found if the topological bit and the probability of existence are utilized. The history of the total cost of the most ﬁt solution in each generation is plotted in Fig. 3.20 for Cases 1 and 2; it shows that the use of the topological bit leads to rapid convergence of the total cost to the optimal value.

128

3.8

Optimization of Finite Dimensional Structures

Random search method using exact reanalysis

3.8.1

Introduction

As we have discussed in previous sections, the cross-sectional areas of frames and trusses are often selected from a catalogue of available values; therefore, a search method that requires a large computational cost for structural analysis should be carried out for ﬁnding the optimal solution with discrete variables. Hence, heuristic approaches can be eﬀectively used for ﬁnding approximate optimal solutions within a practically acceptable computational cost. Among many heuristic approaches, single-point-search methods, e.g., simulated annealing and tabu search, can obtain approximate solutions with a very small number of response evaluations; see Appendix A.3 for details. Since these approaches utilize local search that evaluates the performances of the solutions in the neighborhood of the current solution, the computational cost can be further reduced if the method called exact reanalysis is employed for structural analysis of a slightly modiﬁed structure. In this section, we ﬁrst introduce an exact reanalysis method of trusses subjected to static loads and summarize the random search algorithm in Ohsaki (2001a) for topology optimization of trusses with a discrete list of cross-sectional areas.

3.8.2 3.8.2.1

Exact reanalysis Historical review

The method for computing the exact responses after modiﬁcation of the variable by the speciﬁed amount is called exact reanalysis. There are several approaches to exact reanalysis of static responses of structures, e.g., the expansion method, the reduced basis method, and a method based on inversion of a modiﬁed matrix. Melosh and Luik (1968) presented a simple approach to reanalysis of trusses. Argyris and Roy (1972) presented a more general approach allowing changes in the number of degrees of freedom and support conditions. Kirsch and Rubinstein (1972) discussed convergence properties of the iterative reanalysis process. Kirsch (1994, 1996) and Kirsch and Liu (1995) presented an eﬃcient method combining the reduced basis method and the expansion method. Exact reanalysis of eigenvalues can be found in Chen, Yang, and Lian (2000), Chen, Wu, and Yang (2006), and Kirsch (2000). There are other methods called the virtual distortion method (Putresza and Kolakowski 2001), the pseudoforce method (Deng and Ghosn 2001), and the pseudodistortion method (Makode, Corotis, and Ramirez 1999). Calculation of the inverse of a modiﬁed matrix has been discussed in many ﬁelds of engineering and mathematics; developments in general methodologies are summarized in the review article by Henderson and Searle (1981). Kavlie,

Topology Optimization of Trusses

129

Graham, and Powell (1971) developed a stiﬀness-based method for computing the displacements of the modiﬁed design based on the Sherman-MorrisonWoodbury (SMW) formula (Sherman and Morrison 1950; Woodbury 1950). Akg¨ un, Garcelon, and Haftka (2001) extended the SMW formula to the nonlinear reanalysis problem. However, it is very important in the application to the static analysis problem of structures that the inverse of the stiﬀness matrix of the initial structure is not usually known; i.e. the matrix is only decomposed into triangular matrices (Ohsaki 2001a). 3.8.2.2

Inverse of the modiﬁed matrix

The formulation of the inverse of the modiﬁed matrix using the inverse of the original matrix was studied independently in various areas of mathematics and engineering (Henderson and Searle 1981). In the early 20th century, some formulations were derived indirectly from the inverse of a partitioned matrix. Let K, B, C, and D denote matrices with appropriate size, and I and O be the identity matrix and null matrix, respectively. Schur (1917) derived the following formula: ( )( ) ( ) K−1 O K B I K−1 B = (3.66) −CK−1 I C D O D − CK−1 B from which the relations of the determinants are obtained as K B −1 −1 C D = |K||D − CK B| = |D||K − BD C|

(3.67)

The following formulation of the inverse of a partitioned matrix was explicitly derived by Banachiewicz (1937), although it is sometimes attributed to Schur: ( )−1 K B C D ( −1 ) K + K−1 B(D − CK−1 B)−1 CK−1 −K−1 B(D − CK−1 B)−1 = −(D − CK−1 B)−1 CK−1 (D − CK−1 B)−1 (3.68) Hotelling (1943) derived a similar formula as ( )−1 ( ) K B (K − BD−1 C)−1 −K−1 B(D − CK−1 B)−1 = C D −D−1 V(D − CK−1 B)−1 (D − CK−1 B)−1 (3.69) By comparing the leading terms of (3.68) and (3.69), we obtain the following formula for the inverse of a sum of a matrix: (K − BD−1 C)−1 = K−1 + K−1 B(D − CK−1 B)−1 CK−1

(3.70)

Since the 1950s, more general formulas have been investigated directly for ﬁnding the inverse of a sum of matrices. In particular, if K and D are symmetric and C = B⊤ , the following relation holds (Henderson, Kempthorne,

130

Optimization of Finite Dimensional Structures

Searle, and von Krosigk 1959): (K + BDB⊤ )−1 = K−1 − K−1 B(D−1 + B⊤ K−1 B)−1 B⊤ K−1

(3.71)

which demands D be nonsingular. If D is singular, the following formula can be used (Harville 1976): (K + BDB⊤ )−1 = K−1 − K−1 BD(I + B⊤ K−1 BD)−1 B⊤ K−1 3.8.2.3

(3.72)

Exact reanalysis of trusses

Consider a truss subjected to static nodal loads P. Let n denote the number of degrees of freedom, and K denote the n × n stiﬀness matrix. The nodal displacement vector U0 of the initial (reference) truss before modiﬁcation is obtained from KU0 = P (3.73) The matrix K is a function of the vector A = (A1 , . . . , Am )⊤ of cross-sectional areas, where m is the number of members. Because the components of K are proportional to Ai , K is written as the sum of Ai Ki with Ki being the n × n stiﬀness matrix for unit cross-sectional area of the ith member. Note that the member properties are deﬁned by the matrices and vectors of size n for simple presentation of equations. Computation is carried out, however, using element-size matrices and vectors. Suppose the cross-sectional area of the kth member is increased by ∆Ak . Eq. (3.73) for the modiﬁed truss is formulated as (K + ∆Ak Kk )U1 = P

(3.74)

The objective here is to ﬁnd the displacement vector U1 of the modiﬁed truss for the given load vector P. In the following, the superscripts 0 and 1 denote the values of initial and modiﬁed trusses, respectively. It is important to note here that the inverse of K is not usually known; i.e., K is only decomposed into triangular matrices for computing U0 . Therefore, it is very important to derive a more explicit formula without the inverse of the initial stiﬀness matrix for computing U1 of the modiﬁed truss (Ohsaki 2001a). Let Ni0 denote the axial force of the ith member of the initial truss. The relation between Ni0 and U0 is written by using a vector ci as 0 Ni0 = Ai c⊤ i U

(3.75)

Let E and Li denote the elastic modulus and the length of the ith member, and deﬁne vector bi as Li bi = ci (3.76) E Then the relation between Ni0 and the n-vector of equivalent nodal load vector F0i of the ith member is written as F0i = Ni0 bi . Hence, bi denotes the nodal

Topology Optimization of Trusses

131

load vector corresponding to the unit axial force of the ith member. The matrix Ki is deﬁned by using ci and bi as Ki = bi c ⊤ i

(3.77)

The inverse of the matrix K + ∆Ak Kk is obtained from (Henderson and Searle 1981) ( )−1 −1 (K + ∆Ak Kk ) = K + ∆Ak bk c⊤ k −1 (3.78) ∆Ak K−1 bk c⊤ kK = K−1 − ⊤ −1 1 + ∆Ak ck K bk By post-multiplying P to (3.78), the following relation is derived: U1 = U0 −

0 ∆Ak K−1 bk c⊤ kU −1 b 1 + ∆Ak c⊤ k kK

(3.79)

where (3.73) and (3.74) have been used. Let superscript k∗ denote the values of the initial truss subjected to the load vector bk . Then (3.79) is rewritten as U1 = U0 −

0 ∆Ak Uk∗ c⊤ kU k∗ 1 + ∆Ak c⊤ kU

(3.80)

∆Ak Eε0k U0 =U − 1 + ∆Ak Eεk∗ k 0

where (3.75) has been used, and εk is the strain of the kth member. This way, U1 can be obtained without using K−1 . The strain of the member with a null cross-sectional area may be computed from the displacements of the nodes connected to the member. Therefore, (3.80) can be used for addition or removal of a member between two existing nodes. Next, we consider a case where the cross-sectional areas Ak of members (k = 1, . . . , q) are modiﬁed simultaneously. The stiﬀness matrix of the modiﬁed truss is formulated as K+

q ∑ k=1

∆Ak Kk = K +

q ∑ ∆Ak E k=1

= K + BDB

Lk

bk b⊤ k

(3.81)

⊤

where the ith column of the n × q matrix B is bi , and D is a diagonal matrix deﬁned as D = diag(∆A1 E/L1 , . . . , ∆Aq E/Lq ) (3.82) Then, the inverse of the modiﬁed stiﬀness matrix is written as (3.71). By post-multiplying P on both sides of (3.71), the following relation is derived: U1 = U0 − K−1 B(D−1 + B⊤ K−1 B)−1 DB⊤ U0 = U0 − C∗ (D−1 + B⊤ K−1 B)−1 y

(3.83)

132

Optimization of Finite Dimensional Structures

TABLE 3.2: Results displacements (mm). ua Initial −2.23732 Case 1 −2.11225 Case 2 −2.20397 Case 3 −2.10492

of reanalysis of ub −2.43724 −2.30694 −2.38722 −2.29644

uc −3.05510 −2.90569 −3.02175 −2.89815

where the ith component of the q-vector y is ε0i Li , and the ith column of the n × q matrix C∗ is Ui∗ . A temporary variable vector z is deﬁned as the solution of (D−1 + B⊤ K−1 B)z = y

(3.84)

Eq. (3.84) is simply rewritten as (D−1 + Y∗ )z = y

(3.85)

where the (i, j)-component of Y∗ is equal to εi∗ j Lj ; i.e., the inverse matrix −1 K need not be explicitly computed also for this case. Then U1 is obtained from U1 = U0 − U∗ z (3.86) For the case of the modiﬁcation of a single variable, i.e., q = 1, the ﬁrst component of y is ε01 L1 , and the (1, 1)-components of D and Y∗ are ∆A1 E/L and ε1∗ 1 L1 , respectively. Hence, the ﬁrst component z1 of z is obtained as z1 =

∆A1 Eε01 1 + ∆A1 Eε1∗ 1

(3.87)

Therefore, (3.80) for modiﬁcation of a single variable is a special case of (3.86) for modiﬁcation of multiple variables. Example 3.4 As an example, the nodal displacements after modiﬁcation of cross-sectional areas are computed using exact reanalysis for a 2×2 plane grid truss, as shown in Fig. 3.21. The length of the members in the x- and y-directions are 2.0 m, 2 and P = 100.0 kN. The elastic modulus is 205.8 kN/mm . The displacements ua , ub , and uc in the y-direction of nodes ‘a’, ‘b’, and ‘c’, respectively, for the initial truss with Ai = 1000.0 mm2 for i ̸= 1 and A1 = 0 are as shown in the ﬁrst row of Table 3.2, where members 1, 2, and 3 are deﬁned in Fig. 3.21. Reanalysis has been carried out for the following three cases: Case 1: Increase A2 from 1000.0 mm2 to 1500.0 mm2 and use (3.80) for reanalysis.

Topology Optimization of Trusses

133

a

2

3 b 1

y

c

x

FIGURE 3.21:

P

A 2 × 2 plane grid truss.

Case 2: Increase A1 from 0 to 500.0 mm2 and use (3.80) for reanalysis. Case 3: Increase A2 and A3 to 1500.0 mm2 simultaneously and use (3.85) and (3.86). The results are listed in Table 3.2. It is conﬁrmed that exact responses have been found for all the cases, including Case 2, corresponding to the addition of a member.

3.8.3 3.8.3.1

Random search for topology optimization of trusses General algorithm

We consider truss optimization with inequality constraints for the response b j (A), e.g., nodal displacements and member stresses, which are quantities H given as b j (A) ≤ HjU , (j = 1, · · · , nI ) H (3.88) b j (A), and nI is the number of constraints. where HjU is the upper bound for H Note that the lower-bound constraints are also converted into the standard form and included in (3.88). Then the constraints are normalized assuming HjU > 0 as Hj (A) =

b j (A) H − 1 ≤ 0, (j = 1, . . . , nI ) HjU

(3.89)

The objective function V (A) is the total structural volume. The value of Ai is to be selected from the list A = {a1 , . . . , ar }, where r is the number of

134

Optimization of Finite Dimensional Structures

available sections. Then, the optimization problem is formulated as Minimize

V (A)

(3.90a)

subject to Hj (A) ≤ 0, (j = 1, · · · , n ) Ai ∈ A I

(3.90b) (3.90c)

In order to show the eﬀectiveness of the exact reanalysis method applied to topology optimization of a truss, a random search algorithm is used with a technique for improving convergence to the global optimal solution; see Appendix A.3 for the basic algorithm of random search. The simplest approach of utilizing a random search for a constrained optimization problem may be to add a penalty term of the constraints to the objective function. In this case, however, it is very diﬃcult to assign an appropriate value for the penalty coeﬃcient so that the magnitude of the objective function is equivalent to that of the penalty term. Therefore, the penalty for violating the constraints is implicitly given here by the threshold level for deﬁning the increase or decrease of the design variables. Assuming that Hj (A) is a decreasing function of A, which usually holds for static responses, the optimization algorithm is presented as follows: Step 1: Assign the initial values of A1 , . . . , Am from the given list A, and set parameters H, τ , and d. Step 2: Evaluate the objective function V (A) and the constraints Hj (A), and ﬁnd the maximum value H max among Hj (A) (j = 1, . . . , nI ). At this step, the stiﬀness matrix K is reconstructed and its Cholesky decomposition is carried out. Step 3: Assign the parameter λ = 0.5 + τ H max , where τ (> 0) adjusts the threshold level for increase or decrease of the variable. The parameter λ is replaced with 0 or 1 if λ < 0 or λ > 1, respectively. As is seen in Step 4, λ deﬁnes the probability of increase of the cross-sectional area of a member. If H max < 0, i.e., if all the constraints are satisﬁed, then λ < 0.5 and the cross-sectional areas have a small probability of increase. Step 4: Generate a uniform random number between 0 and 1 to select the member k for which the cross-sectional area is to be modiﬁed. Suppose r allowable values a1 , . . . , ar in A are listed in increasing order, and Ak = ai at the current step. Generate another random number 0 ≤ p < 1 to decide an increase or decrease of Ak . Calculate the smallest integer q, which deﬁnes the amount of increase of Ak , as an integer that is not less than |λ − p|/d with d being the scaling parameter. Modify Ak to aj if p < λ, where j = min{r, i + q}; otherwise decrease Ak to aj with j = max{1, i − q}. Note that a smaller d leads to a larger modiﬁcation of Ak .

Topology Optimization of Trusses

135

1 2 3 4 5 6

7

8

9

10

P FIGURE 3.22:

Optimal solution for the 5 × 5 plane grid.

Step 5: Compute constraint functions for the modiﬁed design using the exact reanalysis method. Accept the design if H max < H, and go to Step 2 if the number of accepted designs is less than the speciﬁed limit. Step 6: Go to Step 4 if H max ≥ H and the number of trials at this stage is less than the speciﬁed limit. Step 7: Select the best design satisfying H max < 0 from the list of accepted designs and terminate the process. 3.8.3.2

Numerical examples

Optimal topologies are found for plane square grids under stress constraints. The ground structure for a 2 × 2 grid is shown in Fig. 3.21. The length of the members in the horizontal and vertical directions are 2.0 m, and P = 2 200.0 kN. The upper bound for the absolute value of stress is 40.0 N/mm . 2 5 U The elastic modulus is 2.058 × 10 N/mm . The parameters are H = 0.3, τ = 10.0, and d = 0.1. In the following, the units of force and length are kN and mm, respectively. The list of available cross-sectional areas is given, including the small value 1.0 as A = {1.0, 2000.0, 4000.0, 6000.0, 8000.0}. Note that the member with Ai = 1.0 is to be removed after optimization and the stress constraint need not be satisﬁed by such a member. An artiﬁcial upper bound U U is given for the nodal displacements to prevent convergence to the obvious and meaningless solution such that Ai = 1.0 for all the members (see Sec. 3.7). The initial solution is given as Ai = 8000.0 for all members. Optimization has been carried out for 2 × 2, 3 × 3, 4 × 4, and 5 × 5 grids starting with 100 diﬀerent random initial solutions for each case. The limit for the number of modiﬁcations of the cross-sectional area is 5000, and the limit for the trial for generating an acceptable neighborhood solution at each step is

136

Optimization of Finite Dimensional Structures

TABLE 3.3:

Result of optimization with 100 diﬀerent random

seeds. UU

save

tave

2 × 2 2.5 3 × 3 5.0 4 × 4 7.0 5 × 5 10.0

469 288 564 1026

8278 4303 7933 14,297

V ave V min V max nopt 8 8 8 (×10 ) (×10 ) (×10 ) 0.072970 0.069292 0.89280 77 1.3094 1.0397 1.7994 26 1.8443 1.3866 2.5491 8 2.5671 1.7227 3.6490 3

2000. Let save and tave denote the average numbers of steps and trial analyses before reaching the minimum objective value within 5000 steps. Average, minimum, and maximum values of the minimum objective value within 5000 steps among 100 diﬀerent random initial solutions are denoted by V ave , V min , and V max , respectively. Table 3.3 shows the values of U U , save , tave , V ave , V min , and V max for each type of grid, where nopt is the number of trials that reached the solution with V = V min ; i.e., the possibility of obtaining the optimal solution decreases as the problem size is increased. It has been conﬁrmed that an appropriate value of U U leads to rapid convergence to nearly optimal solutions. The value of U U is a little greater than the maximum displacement of the truss with optimal topology for each case; i.e., the displacement constraint is not active. The optimal solution of the 5 × 5 grid is shown in Fig. 3.22. The crosssectional areas are 6000.0 for members 6–10 and 8000.0 for the remaining existing members. The value of V min for each case in Table 3.3 corresponds to the objective value of the apparent optimal solution similar to the truss in Fig. 3.22. It may be observed from Table 3.3 that the average number save of analysis that needs decomposition of the stiﬀness matrix is less than 10% of the average number tave of trial analysis that can be done using formulas (3.85) and (3.86) of exact reanalysis. The ratio of CPU time for obtaining optimal topology of the 5 × 5 grid by using the reanalysis method to that without reanalysis is 0.2250. Therefore, computational cost may be drastically reduced utilizing the method of exact reanalysis.

3.9 3.9.1

Multiple eigenvalue constraints Introduction

The eigenvalues of free vibration as well as the linear buckling load factor, which is obtained by solving an eigenvalue problem, are important performance measures of trusses and frames. Therefore, there have been many

Topology Optimization of Trusses

137

studies of optimization under eigenvalue constraints. If the fundamental eigenvalue of the optimum design is simple, the optimization problem may easily be solved using a nonlinear programming or an optimality criteria approach (Venkayya and Tishler 1983; Sadek 1989), because there is no diﬃculty in calculating the sensitivity coeﬃcients of the eigenvalue with respect to the design variable, as presented in Sec. 2.3. However, it is well known that the optimum designs for a speciﬁed fundamental eigenvalue often have multiple (repeated) values. Such an optimal structure was presented by Olhoﬀ and Rasmussen (1977), where necessary conditions for optimality of a column with a variable cross-section are discussed in a variational form, and an optimal column under buckling constraint is found using an optimality criteria approach. Masur (1984) showed that the necessary conditions of Olhoﬀ and Rasmussen (1977) are also suﬃcient conditions in the case of the bimodal optimal solution of symmetric structures. Early developments in this ﬁeld are summarized by Olhoﬀ (1980). Diﬃculties in optimizing distributed parameter structures (continuum) for speciﬁed multiple eigenvalues are discussed extensively in Haug and Cea (1981). It has been shown that multiple eigenvalues are not diﬀerentiable in the ordinary sense, and only directional derivatives or subgradients with respect to the design variables may be calculated. Bochenek and Gajewski (1986) found optimal cross-sectional areas of arches that have at most three-fold eigenvalues corresponding to in-plane and out-of-plane buckling modes. For ﬁnite dimensional structures with a moderately large number of design variables, however, it is very diﬃcult to ﬁnd optimal structures with multiple eigenvalues using conventional optimality criteria approaches or nonlinear programming algorithms. Several computational approaches have been developed for sensitivity analysis of multiple eigenvalues of ﬁnite dimensional structures (Haug and Choi 1982; Haug, Choi, and Komkov 1986; Seyranian 1993) (see Sec. 2.3 for sensitivity analysis of double eigenvalue of vibration). Khot and Kamat (1985) presented an optimality criteria approach for trusses with multiple frequency constraints. Rodorigues, Guedes, and Bendsøe (1995) developed necessary conditions for optimality for problems under constraints on the linear buckling load factor and presented an adjoint variable formulation for sensitivity analysis. Developments in this ﬁeld are summarized in Seyranian, Lund, and Olhoﬀ (1994). In spite of the theoretical developments of sensitivity analysis and optimization methods for problems under multiple eigenvalue constraints, it is very diﬃcult to guarantee global convergence for large-scale structures. Sergeyev and Mr´oz (2000) presented a gradient-based approach that can be applied to the bimodal case and investigated the properties of optimal solution for the problem under stress and frequency constraints. Nakamura and Ohsaki (1988) presented a parametric programming approach to trace a set of optimal solutions under multiple eigenvalue constraints and extended it to truss topology optimization (Nakamura and Ohsaki 1992). Their method has been

138

Optimization of Finite Dimensional Structures

shown to be applicable to topology optimization of frames (Ohsaki and Nakamura 1993). Although their method is eﬀective for the bimodal case, it is very diﬃcult to extend it to obtain optimal solutions with a larger multiplicity of eigenvalues. As an alternative approach, semideﬁnite programming (SDP) has been shown to be eﬀective for structural optimization problems with multiple eigenvalues because it does not need explicit derivatives of eigenvalues (see Appendix A.2.4 for details of SDP). SDP can also be applicable to problems with compliance constraints (de Klerk, Roos, and Terlaky 1995) and robust design (Ben-Tal and Nemirovski 1997). Achtziger and Koˇcvara (2007a, 2007b) presented a nonlinear SDP approach for the problem of minimizing the lowest eigenvalue under constraints on the structural volume and compliance considering multiple load cases. Sequential application of SDP is also eﬀective for the problem with linear buckling constraints (Kanno, Ohsaki, and Katoh 2001).

3.9.2

Multiple eigenvalues in optimal topology

Let Ms (A) and M0 denote the n × n mass matrices corresponding to the structural and nonstructural masses, respectively, where n is the number of degrees of freedom. Note that Ms is a function of the vector A of m design variables. The eigenvalue problem of free vibration is formulated as KΦr = Ωr (M0 + Ms )Φr , (r = 1, . . . , n)

(3.91)

where Ωr and Φr are the rth eigenvalue and eigenmode, respectively, which are implicit functions of A. The vector Φr is ortho-normalized by Φ⊤ r MΦs = δrs , (r, s = 1, . . . , n)

(3.92)

where δrs is the Kronecker delta. The objective here is to ﬁnd an optimal solution for minimizing the total structural volume V (A) under constraint such that the fundamental eigenvalue Ω1 is not less than the speciﬁed lower bound ΩL . Considering the occurrence of multiple eigenvalues, the optimization problem is formulated as

Minimize

V (A) =

m ∑

Ai Li

i=1 L

subject to Ωr (A) ≥ Ω , (r = 1, · · · , n) Ai ≥

ALi

(i = 1, · · · , m)

(3.93a) (3.93b) (3.93c)

where Li is the length of the ith member, and a small positive lower bound ALi is usually given for Ai in order to prevent instability of the truss. The optimal topology is found by removing the members with Ai = ALi as the result of optimization.

Topology Optimization of Trusses L

139

L

(4) (1)

(3) M0

(2)

1 (5)

FIGURE 3.23:

3 H

2

A simple four-bar truss with concentrated mass.

(a) Global mode

FIGURE 3.24:

H

(b) Local mode

Eigenmodes of the four-bar truss.

Example 3.5 Systematic occurrence of multiple eigenvalues is illustrated using a small example of a four-bar truss, as shown in Fig. 3.23, that has a nonstructural mass at the horizontal roller support. The cross-sectional areas of horizontal and vertical members are denoted by A1 and A2 , respectively; i.e., we have two design variables. Suppose, for simplicity, the parameters are given without units as L = 1, H = L/2 = 0.5, and M0 = 1. The elastic modulus is 1, and the mass density of the members is 0.01. If the cross-sectional areas are given as A1 = A2 = 1, which leads to a very small structural mass compared with the nonstructural mass, the eigenvalues are 0.4957, 133.8, and 266.7; i.e., the lowest eigenvalue is much smaller than the second and third eigenvalues. The lowest eigenmode is such that the center node and roller support move horizontally, and the axial deformation of the vertical members is zero, as illustrated in Fig. 3.24(a). This type of mode associated with vibration of nonstructural masses is referred to as the global mode. Suppose A1 = A2 = 1 is given as the initial solution for optimization with lower-bound eigenvalue ΩL = 0.4957. Since the vertical members do not deform in the lowest mode, A2 decreases as the optimization process proceeds, while A1 is almost constant. It is obvious, however, that this pin-jointed truss is unstable if the vertical members are removed. Consequently, A2 should have an extremely small value, and the two lowest eigenvalues coincide even in the

140

Optimization of Finite Dimensional Structures

Eigenvalues

4 3 L

2 1

G

G L

0 0

0.002

0.004

0.006

0.008

0.01

Cross-sectional area

FIGURE 3.25: Relation between the eigenvalues and cross-sectional area A2 of the four-bar truss; G: global mode, L: local mode. optimal solution of this simple truss. One of the lowest eigenmodes is such that the center node vibrates vertically without vibration of the nonstructural mass, as shown in Fig. 3.24(b). This type of mode is referred to as the local mode. The two lowest eigenvalues are plotted in Fig. 3.25 for a range of small values of A2 with A1 = 1, where the solid and dotted lines indicate the lowest and second modes, respectively, and ‘L’ and ‘G’ indicate the local mode and global mode, respectively. From the practical point of view, however, optimal topology with extremely thin vertical members, which are conceived as secondary members, is not needed, and designers are not interested in the local mode that is simply suppressed by adding ﬂexural stiﬀness at the center node.

3.9.3

Semideﬁnite programming for topology optimization

As noted earlier, the convergence of the nonlinear programming algorithm for Problem (3.93) deteriorates if the fundamental eigenvalue becomes multiple. In contrast, the optimal solution with multiple eigenvalues can be found without any diﬃculty by converting Problem (3.93) into an SDP problem (Wolkowicz, Saigal, and Vandenberghe 2000; Ohsaki, Fujisawa, Katoh, and Kanno 1999). As we have shown in Sec. 1.9, the eigenvalue constraint with the prescribed lower bound ΩL is converted to the positive semideﬁniteness of the n × n matrix X deﬁned as X = K − ΩL (Ms + M0 )

(3.94)

For a truss, we have K=

m ∑ i=1

A i Ki , M =

m ∑ i=1

Ai Mi + M0

(3.95)

Topology Optimization of Trusses

141

where Ki and Mi are constant n × n matrices. Let X ≽ O indicate that the symmetric matrix X is positive semideﬁnite. The inner product Y • Z of the n × n matrices Y and Z, for which their (i, j)-components are denoted by Yij and Zij , respectively, is deﬁned as Y•Z=

n ∑ n ∑

Yij Zij

(3.96)

i=1 j=1

Then, using (3.94), (3.95), and X ≽ O, the primal problem of SDP for the topology optimization problem is formulated as Minimize

V (A) =

m ∑

Ai Li

(3.97a)

i=1

subject to

m ∑

(Ki − ΩL Mi )Ai − ΩL M0 ≽ O

(3.97b)

i=1

Ai ≥ ALi , (i = 1, · · · , m)

(3.97c)

where ALi is a small positive lower bound for Ai . The dual of the SDP problem (3.97) is formulated as Maximize ΩL M0 • Y

(3.98a)

subject to (Ki − Ω Mi ) • Y + ηi = Li , (i = 1, . . . , m) ηi ≥ 0, (i = 1, . . . , m)

(3.98b) (3.98c)

L

Y≽O

(3.98d)

where ηi is a slack variable. Notice here that Problems (3.97) and (3.98) are formally categorized as dual and primal forms of SDP problems, respectively. However, we regard Problem (3.97) as a primal problem, because it is derived from the original problem (3.93). Problems (3.97) and (3.98) may be solved simultaneously using the primal-dual interior-point method (Mehrotra 1992). It is important to note here that the sensitivity coeﬃcients of eigenvalues with respect to the design variables A are not needed in the optimization process. Therefore, there is no diﬃculty, as will be shown in the examples, in ﬁnding the solutions with multiple fundamental eigenvalues. Furthermore, it is very important in engineering applications that a symmetric solution is always found using an interior-point method without explicit assignment of symmetry conditions, if the structure is symmetric, and a symmetric initial solution is assigned (Kanno, Ohsaki, Murota, and Katoh 2001; Kanno, Ohsaki, and Katoh 2002). Note that the eigenvalue optimization problem of a frame with a solid section turns out to be a nonlinear SDP problem, because K and/or M are nonlinear functions of the design variables representing the cross-sectional parameters (Kanno and Ohsaki 2007).

142

Optimization of Finite Dimensional Structures

SDP has been shown to be eﬀective for topology optimization of trusses considering compliance under given static loads. The problem under constraints on eigenvalues as well as compliances against multiple loading condiP tions P1 , . . . , Pn is formulated as the following SDP problem (Achtziger and Koˇcvara 2007b): Minimize subject to

V (

) −Pk⊤ ≽ O, (k = 1, . . . , nP ) K(A)

WU −Pk m ∑ Ai Li ≤ V

(3.99a) (3.99b) (3.99c)

i=1

K(A) − ΩL (Ms (A) + M0 ) ≽ O

(3.99d)

Ai ≥ 0, (i = 1, · · · , m)

(3.99e)

where the variables are A and V , and W U is the upper bound for the compliance. Note that (3.95) is to be used for explicit formulation of the constraints with respect to A. Many interior-point methods for linear and quadratic programming have been extended to solve SDPs (Kojima, Shindoh, and Hara 1997; Sturm 1999). Several free libraries, e.g., SDPA (Fujisawa, Kojima, and Nakata 1997), are available. In the following examples, we use SeDuMi Ver. 1.1 (Sturm 1999; P´ olik 2005).

3.9.4

Linear buckling constraint

The linear buckling load factor is a basic performance measure of stability of structures mainly for thin-walled structures and frames with slender members. The multiple buckling load factors of the optimal solution were ﬁrst presented for columns by Olhoﬀ and Rasmussen (1977). Similar to the case of the eigenvalue of vibration, multiple buckling load factors are not continuously diﬀerentiable, and the optimization problem is eﬀectively formulated using an SDP problem with matrix inequality constraints. However, contrary to the vibration problem, the geometrical stiﬀness matrix for formulating the linear buckling problem depends on internal forces that are usually nonlinear functions of the design variables. Therefore, in this section, an optimization method is presented for the nonlinear SDP problem based on the successive solution of linearized SDPs (Kanno, Ohsaki, and Katoh 2001). The optimal cross-sectional areas of trusses subjected to proportional loads are to be found so as to minimize the total volume for the speciﬁed linear buckling load factor. The quasi-static proportional loads are deﬁned by the load factor Λ and the speciﬁed vector P0 of the load pattern as P = ΛP0 . Let KG denote the geometrical stiﬀness matrix corresponding to the unit load factor. The rth linear buckling load factor Λr and the corresponding buckling

Topology Optimization of Trusses

143

mode Φr are deﬁned by (K + Λr KG )Φr = 0, (r = 1, . . . , n)

(3.100)

Note that KG is generally indeﬁnite and there exist positive and negative buckling load factors as the solution of (3.100). However, we are concerned with the smallest positive value of Λr , because the directions of the loads are usually ﬁxed. Therefore, the reciprocals of load factors 1/Λr are numbered in nonincreasing order, i.e., Λ1 is the smallest positive buckling load factor. The positive lower bound ΛL is speciﬁed for positive Λr ; however, no constraint is given for the negative values of Λr , i.e., Λr should satisfy Λr ≥ ΛL or Λr < 0, (r = 1, . . . , n)

(3.101)

which is alternatively written as 1 1 ≤ L , (r = 1, . . . , n) Λr Λ

(3.102)

Hence, the topology optimization problem for minimizing the total structural volume under linear buckling constraints is formulated as Minimize

V (A) =

m ∑

Ai Li

(3.103a)

1 1 ≤ L , (r = 1, . . . , n) Λr (A) Λ

(3.103b)

Ai ≥ ALi , (i = 1, . . . , m)

(3.103c)

i=1

subject to

where ALi is a small positive lower bound for Ai . The generalized eigenvalue problem (3.100) is reformulated as −KG Φr =

1 KΦr , (r = 1, . . . , n) Λr

(3.104)

Using Rayleigh’s principle and the positive deﬁniteness of K, the condition (3.103b) is reduced to Ψ⊤ KG Ψ 1 − ⊤ ≤ L (3.105) Λ Ψ KΨ for any n-vector Ψ; see Appendix A.1.2. Eq. (3.105) is then rewritten as the following condition of positive semideﬁniteness of a matrix: 1 K + KG ≽ O (3.106) ΛL Furthermore, KG is written as follows as a linear function of the vector N(U) = (N1 (U), . . . , Nm (U))⊤ of member forces against P0 , which are functions of nodal displacement vector U: KG =

m ∑ i=1

Ni (U)KGi

(3.107)

144

Optimization of Finite Dimensional Structures

where KGi is an n × n symmetric constant matrix. Consequently, Problem (3.103) can be equivalently reformulated as Minimize

V (A) =

m ∑

Ai Li

(3.108a)

i=1 m m ∑ 1 ∑ subject to L A i Ki + Ni (U)KG i ≽O Λ i=1 i=1 ∑ Ai Ki U = P0

(3.108b) (3.108c)

i=1

Ai ≥ ALi , (i = 1, . . . , m)

(3.108d)

Note that Problem (3.108) cannot be formulated as an SDP problem, because the equality constraints (3.108c), i.e., the equilibrium equations, are nonlinear with respect to A and U for a statically indeterminate structure. Ben-Tal, Jarre, Koˇcvara, Nemirovski, and Zowe (2000) proposed a successive SDP algorithm that solves Problem (3.108) using the semideﬁnite relaxation of (3.108c). Kanno, Ohsaki, and Katoh (2001) proposed a method based on sequential linearization of the nonlinear SDP problem (3.108). Because U is an implicit function of A that is deﬁned by (3.108c), N(U) is also regarded as a function e of A, which is denoted by N(A). Let the superscript (k) indicate the iterei (A) (i = 1, . . . , m) ation counter for the sequential SDP. The axial forces N (k) e (A) deﬁned by the values at A = A(k) using two are approximated by N i methods as e (k) (A) = Ni (A(k) ) + ∇Ni (A(k) )⊤ (A − A(k) ) Method (a) : N i Method (b) :

e (k) (A) N i

= Ni (A(k) )

(3.109a) (3.109b)

In Method (a), Ni is approximated as a linear function of the design variables A utilizing the sensitivity coeﬃcients of N. Method (b) simply ﬁxes Ni at the current solution A(k) . Note that the sensitivity coeﬃcients of buckling load factors are not used in these formulations. Therefore, the optimal design with a large multiplicity of buckling load factors can be found without any diﬃculty. It should be emphasized that a solution obtained by Method (a) satisﬁes the ﬁrst-order optimality conditions of Problem (3.108). Therefore, the solution is guaranteed to be a stationary point of Problem (3.108). On the contrary, the optimality conditions are not generally satisﬁed by the solution obtained using Method (b).

3.9.5

Numerical examples

Optimal topologies are found for plane and space trusses under constraints on eigenvalues of free vibration. The material of the members is steel with

Topology Optimization of Trusses d

FIGURE 3.26:

c

b

145 a

A 6 × 6 plane square grid. 2

the elastic modulus E = 200 kN/mm and the mass density ρ = 7.86 × 10−6 kg/mm3 . SeDuMi Ver. 1.1 (Sturm 1999) is used for optimization. Note that SeDuMi can ﬁnd an optimal solution without explicit assignment of a positive lower bound for Ai . However, the minimum cross-sectional area ALi = 1.0 × 10−6 mm2 is given to prevent the existence of a negative cross-sectional area at the optimal solution due to the nonzero tolerance of the constraints, and also to prevent inaccuracy of the eigenvalue analysis that is carried out for veriﬁcation purposes. The computation was carried out on a PC (Xeon 3.8GHz with 2.0GB memory). Consider ﬁrst a 6 × 6 plane square grid, as shown in Fig. 3.26, where the lengths of the vertical and horizontal members are 200 mm. The optimal topologies are shown in Figs. 3.27(a)–(d), where the concentrated nonstructural mass of 1000.0 kg is located at nodes ‘a’–‘d’, respectively. The speciﬁed eigenvalue is 100.0 rad2 /s2 for all cases. The width of each member in Fig. 3.27 is proportional to its cross-sectional area, and the members with Ai = ALi have been removed. Therefore, the unstable nodes at which two colinear members are connected should be ﬁxed after removal of the thin members. As is seen from Fig. 3.27(d), the symmetric optimal solution has been successfully found, for the case where the mass exists at the center node ‘d’, without explicit assignment of symmetry constraints. Especially in Fig. 3.27(a), a kind of net with small cross-sectional areas should exist to prevent instability of the longer members consisting of large cross-sectional areas. The three lowest eigenvalues, total volume, number of iterations in the primal-dual interior-point algorithm, and CPU time are listed in Table 3.4. For the case with a mass at node ‘a’, the fundamental eigenvalue is duplicated as Ω1 = Ω2 = 100.0 rad2 /s2 , where the corresponding eigenmodes are shown

146

Optimization of Finite Dimensional Structures Eigenvalues (rad2 /s2 ), total volume (×107 mm3 ), number of iterations, and CPU time (s) of the plane square 6 × 6 grid. Location of mass ‘a’ ‘b’ ‘c’ ‘d’ Ω1 (rad2 /s2 ) 100.00 100.00 100.00 100.00 Ω2 (rad2 /s2 ) 100.00 108.08 103.47 100.00 Ω3 (rad2 /s2 ) 115.98 131.12 125.92 102.66 Total volume 8.4811 7.8047 7.4250 7.3065 No. of iterations 29 21 20 22 CPU time 5.28 3.34 3.23 3.91

TABLE 3.4:

(a) Mass at node ‘a’

(b) Mass at node ‘b’

(c) Mass at node ‘c’

(d) Mass at node ‘d’

FIGURE 3.27:

Optimal topologies of the 6 × 6 plane square grid.

in Fig. 3.28. Note that the mode in Fig. 3.28(a) is a global mode in which the node with nonstructural mass vibrates, whereas the mode in Fig. 3.28(b) is a local mode in which the node with mass does not move. As is seen from Table 3.4, the two lowest eigenvalues coincide if the mass exists at node ‘a’ or ‘d’. For the case with a mass at node ‘d’, the two lowest eigenmodes are antisymmetric and symmetric as shown in Fig. 3.29. The number of iterations and the CPU time do not depend on the location of the mass or

Topology Optimization of Trusses

(a) Global mode

147

(b) Local mode

FIGURE 3.28: The two lowest eigenmodes of the optimal 6 × 6 plane square grid with mass at node ‘a’.

(a) Antisymmetric mode

(b) Symmetric mode

FIGURE 3.29: The two lowest eigenmodes of the optimal 6 × 6 grid with mass at node ‘d’. the multiplicity of the eigenvalues. We next ﬁnd optimal solutions for double-layer cylindrical grid trusses with diﬀerent numbers of grids, where the conﬁguration of the 7 × 7 grid is shown in Fig. 3.30. The truss has pin supports at the four lower corners and the lumped mass of 1000 kg at each node. Let nG denote the number of units in both the longitudinal and span directions of the upper surface. The span length (m) between the supports is 2(nG − 1) for the two directions. The open angle of the lower cylinder is 80 deg., and the distance between the lower and upper cylinders is 2 m. The lower and upper chords in each direction, and the diagonals have the same lengths, respectively. The speciﬁed fundamental eigenvalue is 1000 rad2 /s2 . The optimal cross-sectional areas are shown in Figs. 3.31(a)–(d). As is seen, the lower chords and the diagonals near the supports as well as the upper chords near the center lines have large cross-sectional areas. The three eigenvalues coincide at the optimal solutions, as shown in the second column in Table 3.5, and the symmetry of the cross-sectional areas is attained without

148

Optimization of Finite Dimensional Structures

z y x

FIGURE 3.30:

A 7 × 7 double-layer cylindrical grid.

Eigenvalues (rad2 /s2 ), total volume (m3 ), number of iterations, and CPU time (s) of the cylindrical double-layer grid.

TABLE 3.5:

nG n m Ω1 Ω2 Ω3 Ω4 Total volume No. of iterations CPU time CPU/(n + m)3 × 107

7 327 392 1000.0 1000.0 1000.0 1134.0 1.9259 32 163.3 4.393

8 423 512 1000.0 1000.0 1000.0 1217.1 4.2570 35 449.0 5.493

9 531 648 1000.0 1000.0 1000.0 1287.2 8.9832 34 904.7 5.520

10 651 800 1000.0 1000.0 1000.0 1350.7 18.401 37 1829 5.987

assignment of explicit symmetry constraints. The three modes are as shown in Fig. 3.32. Mode 1 is symmetric with respect to the xz- and yz-planes, while Modes 2 and 3 are antisymmetric with respect to the xy- and yz-planes, respectively. Optimization was also carried out for 8 × 8, 9 × 9, and 10 × 10 grids. The four lowest eigenvalues, total volume, number of iterations, and CPU time are listed in Table 3.5. Note that the algorithm did not converge for the 10 × 10 grid, and stopped with numerical diﬃculty; however, a feasible and nearly optimal solution was successfully found, because a primal-dual interior-point method that searches in the feasible region was used. As is seen from the last row of Table 3.5, the CPU time is almost proportional to (n + m)3 , as it is known that the computational cost of the interior-point method can be

Topology Optimization of Trusses

149

y

x

(a) Diagonal view

(b) Lower chords

(c) Upper chords

(d) Diagonals

FIGURE 3.31: Optimal solution of the 7 × 7 double-layer cylindrical grid.

estimated as a polynomial of the problem size.

3.10 3.10.1

Application of data mining Frequent item set of decent solutions

Topology optimization may be conceived as a process of selecting the best solution from the set of the huge number of feasible solutions. For example, it is possible to select the optimal topology from the candidate list of trusses, as shown in Fig. 3.33, that represent substructures of a ground structure consisting of 10 bars that can exist. However, it is not possible to enumerate

150

Optimization of Finite Dimensional Structures

(a) Mode 1

(b) Mode 2

(c) Mode 3 FIGURE 3.32: Three fundamental eigenmodes of the optimal solution of the 7 × 7 double-layer cylindrical grid.

all stable topologies from a ground structure with moderately large numbers of nodes and members. Therefore, methodologies developed in the ﬁeld of data mining can be used for ﬁnding the set of existing/nonexisting members that leads to global optimal topology. Data mining consists of the methods of association rule, decision tree, clustering, etc., for extracting the rules and knowledge from a huge database (Hand, Mannila, and Smyth 2001; Berry and Linoﬀ 1997). The basic approaches of data mining for application to topology optimization are illustrated using a 10-bar truss in Fig. 3.34. Loads of 0.02 and −0.1 are applied in the x- and y-directions, respectively, at all the nodes, where units are omitted for brevity. The elastic modulus is 1, and the cross-sectional area of each member takes 0 or 1, for simplicity; i.e., the existing members have the same cross-sectional area, and, consequently, the total number of possible combinations of the members is 210 . We consider a multiobjective optimization problem for minimizing the total structural volume and compliance. All of the possible 210 topologies have been enumerated to ﬁnd only 40 stable topologies. The values of total structural volume and compliance of the 40 stable solutions are classiﬁed into four clusters in the objective function space using the data mining tool WEKA Ver. 3.5 (Witten and Frank 2000). The solutions in the four clusters are plotted with diﬀerent marks in Fig. 3.35. The well-known method of K-means has been used for clustering, which is described as follows:

Topology Optimization of Trusses

(2)

(3)

(4)

(5)

(6)

(7)

(8)

An example of a list of candidate topologies. (1)

(3)

1

(5)

2

5 8 y

6 10

3 (2) x

FIGURE 3.34:

7

9 4

(4) 2.0 m

2.0 m

FIGURE 3.33:

(1)

151

(6) 2.0 m

A 10-bar truss.

Step 1: Select four solutions that are preferably widely spaced in the objective function space, each of which is considered to be located at the center of each cluster. Step 2: Assign each solution to the cluster that has the nearest center. Step 3: Compute the center of each cluster in the objective function space, and go to Step 2 if not converged. The 17 solutions indicated by ‘+’ in Fig. 3.35 form a cluster of decent solutions that have smaller values of compliance and total structural volume.

152

Optimization of Finite Dimensional Structures 8

Compliance

6 4 2 0 16

18

20

22

24

Total volume

FIGURE 3.35: Clustering of the stable solutions of the 10-bar truss in the objective function space.

Each solution is designated by the set of existing members. Therefore, using the terminologies of data mining, the member and solution correspond to item and transaction, respectively. Let d denote the number of elements in the transaction database D. In this example, D consists of 40 stable solutions. Let Gi denote the cover of the ith item (member) in D; i.e., Gi is the set consisting of the solutions including member i. The number of elements (solutions) in Gi is denote by gi . Then the frequency fi of item (member) i is deﬁned as fi = gi /d, and the set of items that has a frequency greater than the speciﬁed value α is called the frequent item set. Next we deﬁne the set Z of the decent solutions that have a particular favorable property, e.g., small values of compliance and total structural volume for the 10-bar truss, and deﬁne the set Hi of solutions that have member i in Z. The number of solutions in Hi is denoted by hi . Consider a rule Gi → Z, which states that the solution is a decent solution if it has member i. The support si and conﬁdence ci of this rule are deﬁned by si = hi /d and ci = hi /gi , respectively. This way, a reliable rule is characterized by large support and conﬁdence. Since the loads are applied at all nodes of the 10-bar truss, all nodes should exist so that the truss is stable. Therefore, the support of any rule Gi → Z has a moderately large value that is greater than 0.386. The conﬁdences ci of the rule Gi → Z are c1 = c3 = 0.515, c2 = c4 = c6 = c9 = c10 = 0.412, and c5 = c7 = c8 = 0.333. Therefore, the stable solution with member 1 or 3 has a high possibility of being a decent solution. In contrast, let G i denote the set of solutions that do not include member i. The set of solutions that are not included in the set of decent solutions is denoted by Z. For the rule G i → Z, the members 1 and 3 have support 0.175 and conﬁdence 1, which agrees with the results obtained by intuition; i.e., a truss has a large compliance if it does not have the horizontal members connected to the supports.

Topology Optimization of Trusses

153

5

6

5

6

5

6

3

4

3

4

3

4

1

2

1

2 (c) 2 bars

(b) 2 bars

1

2

5

6

5

6

3

4

3

4

1

2

1

2

(a) fully braced frame

(d) 3 bars

(e) 4 bars

FIGURE 3.36: Predeﬁned fully braced frame and ground structures as its subsets.

3.10.2

Topology mining of ground structures

One of the main diﬃculties in topology optimization of trusses and braces under stress constraints is that the local optimal solution obtained using an NLP algorithm strongly depends on the initial ground structure (GS), as demonstrated in Example 3.2, and the solution does not always improve as a result of increasing the number of members in the GS. Therefore, the methodologies developed in the ﬁeld of data mining can be used for ﬁnding the optimal ground structure that has a moderately small number of members and leads to the global optimal topology. Hagish*ta and Ohsaki (2008b, 2008b) presented a method called topology mining for ﬁnding GSs that lead to the global optimal topology, where the superior bar sets are extracted using the apriori algorithm developed by Agrawal, Imielinski, and Swami (1998). For example, consider a problem of optimizing the locations of the braces of a frame with ﬁxed cross-sectional areas, such as the fully braced frame shown in Fig. 3.36(a). The total structural volume is minimized under stress constraints. The GSs shown in Figs. 3.36(b)–(e) are obtained as subsets of the fully braced frame. We assume that singular optimal topologies may exist for this problem; i.e., the stress constraints may not be satisﬁed by nonexistent braces.

154

Optimization of Finite Dimensional Structures

TABLE 3.6:

Expression of topologies of the ground structures. Ground structure Optimal objective value (b) {114, 123, 036, 045} Vb (c) {014, 123, 136, 045} Vc (d) {114, 123, 036, 145} Vd (e) {114, 123, 136, 145} Ve

5

6

5

6

3

4

3

4

1

2

1

2

: frequent bars

: infrequent bars

FIGURE 3.37: Superior bar sets consisting of three braces in ground structures in Fig. 3.36(b)–(e).

The topologies of the braced frames in Figs. 3.36(b)–(e) are expressed by integer arrays, as shown in Table 3.6, where the existing bar between nodes i and j is represented by the three digits with ‘1’ followed by the node numbers ‘i’ and ‘j’ as ‘1ij’, while the nonexistent bar between nodes j and k is indicated by 0jk with ‘0’ at the ﬁrst digit. The GS deﬁned by each array in the table is optimized using an NLP algorithm to ﬁnd the optimal objective value denoted by Vb , . . . , Ve . Let β denote the speciﬁed ratio of the number of superior ground structures to the total number of GSs. If β = 0.25 for 16 (= 24 ) GSs for the frames with four braces, then the four best GSs are regarded to be superior. Suppose the frames in Figs. 3.36(b)–(e) are the four superior ground structures. The frequent bar set is deﬁned as the set of items (braces) that has large frequency in the superior GSs. The infrequent bar set is also deﬁned as the set of large frequency of nonexistence in the superior GSs. Note that infrequent bar does not mean a bar that exists in nonsuperior GSs. Then, we can obtain the superior bar sets, including the frequent and infrequent bar sets, by applying the apriori algorithm to the ground structures in Table 3.6. The minimum frequency is assigned as 0.5; i.e., the set of braces that are existent/nonexistent in two or more ground structures is extracted. First, we can see from Figs. 3.36(b)–(e) that the frequent/infrequent bar sets, which

Topology Optimization of Trusses 2

4

155

6 1.0

1 1.0

1.0 (a)

V = 13.505 (b)

7 P2

5 P1

3

1.0

(c)

(d)

(e)

(f)

V = 13.400 (g)

FIGURE 3.38: Optimization results of truss topology optimization: (a) fully connected ground structure, (b) optimal 11-bar truss from the fully connected ground structure, (c)–(f) ground structures that lead to the optimal 10-bar truss, (g) optimal 10-bar truss. are called superior bar sets for brevity, consisting of single brace are given as 114, 123, 136, 145, 036, 045

(3.110)

Then, utilizing the apriori algorithm, we should check the following sets to obtain the superior bar sets consisting of two braces: {114, 123}, {114, 136}, {114, 145}, {114, 036}, {114, 045}, {123, 136}, {123, 145}, {123, 036}, {123, 045}, {136, 145}, {136, 045}, {145, 036}, {036, 045}

(3.111)

Note that a brace cannot be existent and nonexistent simultaneously; hence, sets such as {136, 036} have been excluded. For example, the braces connecting nodes 1 and 4, and 2 and 3, respectively, exist simultaneously in three ground structures in Figs. 3.36(b)–(e); however, the braces connecting nodes

156

Optimization of Finite Dimensional Structures

: frequent bar

FIGURE 3.39:

: infrequent bar

An example of superior bar sets.

3 and 6, and 4 and 5, respectively, exist simultaneously only in one ground structure. By checking all the sets in (3.111), we have the following superior bar sets consisting of two braces: {114, 123}, {114, 145}, {114, 036}, {123, 136}, {123, 145}, {123, 036}, {123, 045}

(3.112)

The superior bar sets consisting of three braces are the subsets of the union of the two sets in (3.112). Therefore, we should check {114, 123, 145}, {114, 123, 136}, {114, 123, 036}, {114, 123, 045}, {123, 136, 145}, {123, 136, 045}

(3.113)

From (3.113), the following superior bar sets are obtained: {114, 123, 145}, {114, 123, 036}

(3.114)

which are illustrated in Fig. 3.37. This way, the ground structures that contain superior bar sets can be obtained with moderately small computational cost. Utilizing this approach, Hagish*ta and Ohsaki (2008b) developed a method called topology mining for generating optimal ground structures that leads to global optimal topology. As an example, optimal ground structures for a plane truss are found using topology mining. The fully connected ground structure has 16 members, as shown in Fig. 3.38(a). The truss is subjected to two loads, P1 and P2 , independently at nodes 5 and 7, respectively; see Hagish*ta and Ohsaki (2008b) for details. A simple application of the ground structure approach to the fully connected ground structure in Fig. 3.38(a) generates the 11-bar truss in Fig. 3.38(b), where V = 13.505. In order to generate the superior bar sets, ground structures are randomly generated 105 times and the optimal topologies are found

Topology Optimization of Trusses

157

from the 11,047 stable ground structures. As a result, we found four ground structures, shown in Figs. 3.38(c)–(f), that lead to the optimal 10-bar truss in Fig. 3.38(g) with the total structural volume V = 13.400, which is less than that of the 11-bar truss in Fig. 3.38(b). Topology mining is then carried out to ﬁnd the superior bar sets, as shown in Fig. 3.39, which lead to the optimal 10-bar truss in Fig. 3.38(g). Note that these sets have been found with application of NLP for 140 stable solutions, which is very small compared with 11,047 for random enumeration.

Chapter 4 Configuration Optimization of Trusses

Conﬁguration optimization of trusses is categorized into topology optimization and geometry optimization, as described in Secs. 1.1 and 1.10, and so far we have mainly considered topology optimization. In this chapter, some methods are presented for conﬁguration optimization of trusses for optimizing topology, nodal locations, and cross-sectional areas simultaneously. A historical review is presented in Sec. 4.1 as an introduction to this chapter. Basic formulations and small examples are given in Sec. 4.2. An optimization method is presented for regular plane trusses with uniform cross-sectional area in Sec. 4.3. Link mechanisms are generated using the extended ground structure approach considering geometrical nonlinearity in Sec. 4.4.

4.1

Introduction

Geometry optimization of trusses and frames seems to be rather straightforward, because the nodal coordinates are considered as continuous design variables, and the optimal solutions may be found using appropriate methods of mathematical programming (Sadek 1986; Imai and Schmit 1982; Dems and Gatkowski 1995; Lin, Che, and Yu 1982; Saka 1980; Svanberg 1981). Early works on geometry optimization are found in Dobbs and Felton (1969) and Pedersen (1972). In the process of geometry optimization, however, side constraints are usually given for the nodal coordinates to prevent unfavorable numerical instability due to the existence of extremely short members or closely spaced parallel members (Sadek 1986; Imai and Schmit 1982). Intersection of members should also be prevented in some cases. Therefore, it is very diﬃcult to modify the topology of a truss or a frame through the process of geometrical optimization by continuously varying the nodes and removing the coalescent nodes and members. For these reasons, it is widely recognized that simultaneous optimization of topology and geometry of a truss is very diﬃcult. It is important to note that the topology optimization problem is a combinatorial optimization problem, whereas the geometry optimization problem

159

160

Optimization of Finite Dimensional Structures

has continuous variables and functions. Therefore, a hybrid approach may be used in which geometry optimization problems are solved successively for the truss obtained at each step of the upper-level topology optimization process (Hagish*ta and Ohsaki 2008b). However, computational eﬀort seems to be very large if such two-level algorithms are applied to optimize large-scale trusses. Achtziger (2007) presented a mathematical programming approach based on implicit programming (Ben-Tal, Koˇcvara, and Zowe 1993). Contrary to the ground structure approach to topology optimization, which allows only removal of members, some methods have been proposed for adding nodes and members from a simple base truss to generate an optimal topology and geometry (Rule 1994; McKeown 1998). The drawback of obtaining unrealistic or unstable optimal solutions for the ground structure approach may be avoided if a growing process is used. There is no theoretically clear criteria, however, for addition of nodes and members (Kirsch 1989a, 1996). The modiﬁcation process of topology may be written in a formal shape grammar (Reddy and Cagan 1995a, 1995b). Shea, Cagan, and Fenves (1997) and Shea and Smith (2006) optimized transmission towers using shape annealing. Bojczuk and Mr´oz (1998b, 1999) used sensitivity information to determine whether the candidate modiﬁed topology is acceptable, and topology is modiﬁed by dividing a bar, or by adding a node in a triangular unit. Rychter and Musiuk (2007) deﬁned the topological sensitivity of eigenvalues with respect to the ﬂip of a diagonal member of a regular truss. Hagish*ta and Ohsaki (2009) presented a growing and removal process based on the mechanical properties of the modiﬁed structure. Yang and Soh (2002) used genetic programming for evolutionary optimization. Stadler (1999) developed an algorithm using the Steiner points of triangular units. Mr´oz and Bojczuk (2003) presented a method similar to a greedy method for ﬁnding the most eﬃcient topology variation of a truss under compliance constraint.

4.2

General formulation and methodologies of configuration optimization

The problem of conﬁguration optimization can be formulated using a socalled extended ground structure approach, where both the cross-sectional areas and nodal coordinates are considered as design variables (Koˇcvara and Zowe 1995). As noted in Sec. 4.1, the main diﬃculty of this problem is that some nodes may coincide during optimization, and the stiﬀness matrix becomes singular due to the existence of a member with zero length. Although these coalescent nodes (Ohsaki 1997b) or melting nodes (Achtziger 2007) should be avoided from a computational point of view, they are preferred because solutions with various topologies and geometries can be found

Conﬁguration Optimization of Trusses (1)

1

(3)

6

(5)

5 3 y

4

9 7

(4)

(6) P

x W

FIGURE 4.1:

H

10 8

2 (2)

161

W

A 10-bar truss with variable nodal locations.

from the ground structure with relatively small numbers of members and nodes. Let A = (A1 , . . . , Am )⊤ denote the vector of cross-sectional areas, where m is the number of members in the initial ground structure. The vector consisting of variable nodal coordinates is denoted by X = (X1 , . . . , XnC )⊤ , where XnC is the number of variables of nodal coordinates. Note that X includes the coordinates in two directions for a plane truss and three directions for a spatial truss. The length of the ith member is denoted by Li (X). The lower bound for Ai is 0 in order to allow topology variation, while the upper bound is not given. If the upper bound XiU and the lower bound XiL are appropriately assigned for Xi to prevent the existence of coalescent nodes, the optimization problem for minimizing the total structural volume V (A, X) under general inequality constraints Hj (A, X) ≤ 0 (j = 1, . . . , nI ) representing the constraints on stresses, displacements, etc., is formulated as Minimize

V (A, X) =

m ∑

Ai Li

(4.1a)

i=1

subject to Hj (A, X) ≤ 0, (j = 1, . . . , nI ) ≤ Xi ≤ (i = 1, . . . , n ) Ai ≥ 0, (i = 1, . . . , m) XiL

XiU ,

C

(4.1b) (4.1c) (4.1d)

Example 4.1 As an illustrative example, consider a 10-bar truss, as shown in Fig. 4.1, subjected to a vertical load P , where W = H = 2 m. Optimal nodal locations are found for two cases with uniform and non-uniform cross-sectional areas, respectively. The total structural volume is minimized under constraint on the compliance. The locations of the loaded node and the supports are ﬁxed. The gray regions in Fig. 4.1 are the feasible regions of the nodes with variable

162

Optimization of Finite Dimensional Structures

(a) Uniform cross-sectional area

(b) Independent cross-sectional area

FIGURE 4.2: Optimal cross-sectional areas and nodal locations of the 10bar truss. coordinates, where the bounds of the coordinates are ±0.5 m from the initial values in the x- and y-directions. The parameters are the same as those for the 10-bar truss in Example 1.10 in Sec. 1.8. SNOPT Ver. 7.2 (Gill, Murray, and Saunders 2002), which utilizes sequential quadratic programming is used for optimization. If the cross-sectional areas of all members have the same value, which is adjusted to have the speciﬁed compliance, the optimal shape is as shown in Fig. 4.2(a). Note that nodes 3, 4, and 5 exist, respectively, in the interior, along an edge, and at a vertex of the feasible region. If the cross-sectional areas of all members can vary independently, the optimal shape and crosssectional areas are as shown in Fig. 4.2(b). Since small lower bounds are given for the cross-sectional areas, shape and topology can be simultaneously optimized by removing those members with lower-bound cross-sectional areas. However, the shape variation is limited within the feasible region of the nodal coordinates. As an alternative formulation, a small lower bound LLi is given for Li (X) to prevent the existence of coalescent nodes as Minimize

V (A, X) =

m ∑

Ai Li

(4.2a)

i=1

subject to Hj (A, X) ≤ 0, (j = 1, . . . , nI )

(4.2b)

Li (X) ≥ LLi , (i = 1, . . . , m)

(4.2c)

Ai ≥ 0, (i = 1, . . . , m)

(4.2d)

Problems (4.1) and (4.2) are standard nonlinear programming problems. However, it is not desirable to directly solve these problems for general largescale trusses, because they are highly nonlinear, and the optimization procedure often terminates at a local optimal solution. One possible approach for improving the convergence to an approximate optimal solution is to ﬁx the

Conﬁguration Optimization of Trusses

163

sizing variables A and the shape variables X alternatively, and ﬁnd the optimal values of the free variables. Pedersen (1972) presented a method based on successive linearization. Consider a case of a single loading condition, where the truss is subjected to a set of nodal loads P = (P1 , . . . , Pn )⊤ with n being the number of degrees of freedom. Achtziger (2007) presented a method based on the implicit programming approach (Ben-Tal, Koˇcvara, and Zowe 1993). Let X = (X1 , . . . , XnC )⊤ denote the vector of variable nodal coordinates. Although the details are omitted here, the conﬁguration optimization problem of minimizing the compliance under constraint on total structural volume is reformulated as Minimize subject to

m ∑

(λi + µi )X⊤ Ci X

(4.3a)

√ (µi − λi ) Ei QCi X + P

(4.3b)

λi ≥ 0, µi ≥ 0, (i = 1, . . . , m)

(4.3c)

i=1 m ∑ i=1

where λi , µi (i = 1, . . . , m), and X are the variables, Ci (i = 1, . . . , m) and Q are nC × nC constant matrices, and Ei is the elastic modulus of the ith member. As is seen, the objective and constraint functions are polynomials of the variables and the maximum order is three in the objective function. Alternatively, an approximate solution can be obtained by modeling the truss using the frame element to prevent diﬃculties due to the existence of coalescent nodes. In the following example, we ﬁrst verify that the displacements of trusses can be approximated by a frame member with an artiﬁcially small radius of gyration of area. Example 4.2 Consider again the 10-bar truss in Fig. 4.1. Suppose the members have sandwich sections such that Ii = h2 Ai /4, where Ii is the second moment of inertia of the ith member, and h is the distance between the ﬂanges that is the same for all members. Other parameters are the same as those in Example 1.3 in Sec. 1.5. The optimal nodal locations and cross-sectional areas are found using SNOPT Ver. 7.2 (Gill, Murray, and Saunders 2002), where a very strict tolerance, 10−8 , is given for constraints and optimality conditions. First, the existence of coalescent nodes is not allowed, and the optimal solution is found under the same conditions as the truss in Fig. 4.1. Then, for h = 1, 10, and 100 (mm) the same topology and almost the same geometry as Fig. 4.2(b) have been obtained. Table 4.1 shows the optimization results for various values of h as well as the result of the truss. As is seen, the optimal cross-sectional areas and the total structural volume converge to those of the truss as h is decreased. Although the bending deformation may be very large if h is small, the average axial deformation can be successfully approximated by the frame model, because

164

Optimization of Finite Dimensional Structures

TABLE 4.1: Optimal cross-sectional areas and total structural volume of the 10-bar truss modeled by truss and frame elements with various values of h (mm). Ai (mm2 ) Truss h = 1 h = 10 h = 100 1 1383.54 1383.54 1383.53 1382.85 2 866.85 866.85 866.83 865.22 3 798.41 798.41 798.41 799.19 4 281.02 281.02 281.02 280.79 5 0.100 0.100 0.100 0.100 6 0.100 0.100 0.100 0.100 7 683.78 683.78 683.77 682.87 8 0.100 0.100 0.100 0.100 9 958.65 958.65 958.65 958.37 10 0.100 0.100 0.100 0.100 Total volume (×107 mm3 ) 1.43036 1.43036 1.43034 1.42875 Member number

FIGURE 4.3: Optimal conﬁguration and cross-sectional areas obtained by the frame model.

we consider only the inﬁnitesimal deformation and each member consists of single element. Optimal solutions are next found for the 10-bar truss without side constraints on the nodal locations. However, a small lower bound LL is given for the length of each member to prevent the existence of too short members that leads to singularity of the stiﬀness matrix. The optimal conﬁguration and cross-sectional areas obtained from the 10-bar truss with H = 4.0 m and W = 2.0 m in Fig. 4.1 are shown in Fig. 4.3, where (h, LL ) = (1, 1). Note that the members with Ai = ALi have been removed in Fig. 4.3. As is seen, a frame with four members and four nodes, including the supports, has been successfully generated. Table 4.2 shows the optimal nodal locations for various values of (h, LL ). Note that the optimization process does not converge strictly if LL = 0.0001 mm. It is seen from Fig. 4.1 and Table 4.2 that nodes 4

Conﬁguration Optimization of Trusses

165

TABLE 4.2:

Optimal nodal locations and total structural volume of the 10-bar truss modeled by frame elements. Node

Direction

3

x y 4 x y 5 x y Total volume (×106 mm3 )

(1,1) 3203.19 1756.53 3999.70 −0.953174 3205.39 1756.43 6.97819

(a) Separation of a node FIGURE 4.4:

(h, LL ) (1,0.1) (0.1,1) (0.0001,1) 3205.81 3205.38 3210.65 1753.85 1754.64 1749.05 4000.06 3998.92 3999.00 −0.0895076 −0.170447 −0.0844409 3519.21 3385.73 3386.83 2839.75 5228.31 1133.13 6.96465 6.96896 6.97272

(b) Addition of a node

Topology variation of a truss.

and 6 are coalescent. Because the cross-sectional areas of members connected to node 5 are very small, the location of node 5 does not have any eﬀect on the optimal solution and its objective value. Finally, optimal nodal locations and cross-sectional areas are found for the truss that has the same topology as the frame in Fig. 4.3. The optimal objective value is 6.961021 × 106 mm3 , which is very close to the values in Table 4.2 obtained using the frame model. Therefore, optimal conﬁguration of a truss can be successfully found after obtaining the optimal topology of the frame model. So far, we have discussed nonlinear programming approaches for conﬁguration optimization of trusses. Several heuristic approaches have been proposed utilizing topological variation in the optimization process. By allowing the addition of nodes and members, as well as variation of nodal coordinates, a complex optimal solution can be found starting with a simple ground structure (Bojczuk and Mr´oz 1998b). Examples of topology variation are illustrated in Fig. 4.4. The location of

166

Optimization of Finite Dimensional Structures

a new node that most eﬃciently reduces the objective function is estimated using the sensitivity coeﬃcients of the objective function with respect to the coordinates of the added or separated node. After addition of a node and the corresponding members, optimization of nodal location is carried out, and another node is to be added. A node may also be added at the intersection of the members (Mart´ınez, Mart´ı, and Querin 2007). Reddy and Cagan (1995b) presented an addition scheme of a triangular unit. One of the main diﬃculties for application of the (extended) ground structure approach to conﬁguration optimization is that a ground structure with many nodes and members is needed to obtain the optimal conﬁguration with moderate complexity. However, in this case, optimization using nonlinear programming often reaches a local optimal solution especially for a problem with stress constraints. Therefore, a kind of optimal ground structure that can reach the global optimal solution from suﬃciently small numbers of nodes and members should be found. For this purpose, Hagish*ta and Ohsaki (2009) presented an approach based on sequential topology optimization for the updated ground structure. In order to search solutions in wider design space, it is important that the new ground structure should be deﬁned to increase the complexity of the design space. The algorithm is summarized as Step 1: Deﬁne the initial ground structure that has suﬃciently small numbers of nodes and members. Step 2: Find the optimal topology considering the cross-sectional areas as design variables and remove the unnecessary nodes and members. Step 3: Add nodes and members to the optimal topology to update the ground structure so as to have maximum disturbance to the optimal topology.

4.3 4.3.1

Optimization of a regular grid truss Problem formulation

In this section, we summarize the diﬃculties in conﬁguration optimization of trusses and present a systematic approach to optimize a regular truss (Ohsaki 1997b). Because the trusses in practical applications in civil engineering are usually regular, and the cross-sectional areas of the members are not allowed to have arbitrarily diﬀerent positive values, it is practically desired to develop an optimization method for such trusses. Consider a pin-jointed regular plane truss, as shown in Fig. 4.5, that consists of rectangular units with ﬁxed topology. The size of a unit is allowed to vary under the condition that the regularity of the truss is preserved. The

Conﬁguration Optimization of Trusses

167

W Y3 Y2 H Y1

y

X0

x X1

X2

X3

X4

X5

Y0

X6 Pa

FIGURE 4.5:

A plane rectangular truss (Nxu = 6, Nyu = 3).

numbers of units in the x- and y-directions are denoted, respectively, by Nxu and Nyu . The cross-sectional areas are assumed to be the same for all members. The problem considered here is to ﬁnd the optimal topology, optimal nodal locations, and optimal cross-sectional area of a regular plane truss. The locations of the supports as well as the geometrical properties, e.g., the span length for deﬁning the total dimension of the truss, are ﬁxed in order to prevent the existence of an impractical optimal solution. The loads are applied at the limited number of nodes for which the locations are ﬁxed. Let n denote the number of degrees of freedom of the initial ground structure. The displacement vector is denoted by U = (U1 , . . . , Un )⊤ , and constraints are given as |Uj | ≤ UjU , (j = 1, . . . , n)

(4.4)

where UjU is the upper bound for the absolute value of the jth displacement component. The design variables are the vector of variable nodal coordinates X and the cross-sectional area A that is the same for all the members. Note that the sizes of X and U depend on Nxu and Nyu . The set of vectors of nodal coordinates corresponding to feasible regular geometry is denoted by X . In the process of designing trusses, the cost of nodes is usually equivalent to that of the members. Therefore, from a practical point of view, it is very important to consider the nodal cost in a topology optimization problem as presented in Sec. 3.7.3. The cost of each node is denoted by B, which is the same for all the nodes. Let N n denote the total number of nodes, which depends on Nxu and Nyu , and denote by cm the cost coeﬃcient for the unit volume of members. The length of the ith member is denoted by Li , which is a function of X. Then the conﬁguration optimization problem of minimizing

168

Optimization of Finite Dimensional Structures

the total cost C(X, A) for ﬁxed Nxu and Nyu is stated as Minimize

C(X, A) = cm

subject to |Uj (X, A)| ≤

m ∑

i=1 UjU ,

ALi (X) + N n B

(4.5a)

(j = 1, . . . , n)

(4.5b)

X∈X

(4.5c)

Example 4.3 To illustrate the diﬃculties in conﬁguration optimization of a truss with uniform cross-sectional area, optimum designs are found for a simple cantilevertype plane truss with ﬁxed span length W and height H. Nodal cost is not considered here for a simple presentation of the results. A truss corresponding to (Nxu , Nyu ) = (3, 1) is shown in Fig. 4.6. The vertical member is assumed to remain vertical during design modiﬁcation; i.e., the feasible trusses consist of rectangular units. The x-coordinates Xi (0 ≤ i ≤ Nxu ) of the boundaries of the units are deﬁned as shown in Fig. 4.6. The coordinates Xi (1 ≤ i ≤ Nxu − 1) and the member cross-sectional area A are chosen as independent design variables. The load Pa = 49.0 kN is applied in the y-direction at node ‘a’. The parameters are U U = 10.0 mm, cm = 1.0 × 10−6 mm−3 , W = 4000.0 mm, 2 H = 2000.0 mm, and the elastic modulus is E = 205.8 kN/mm . In the following, the units of force and length are kN and mm, respectively, and are omitted for brevity. Optimal trusses are found for Nxu = 1, 2, 3, and 4 using the optimization package IDESIGN (Arora and Tseng 1987), which utilizes sequential quadratic programming. The values of Xi , A, and C of the optimum designs with ﬁxed topologies are listed in Table 4.3. It is interesting to note that equally divided conﬁgurations are optimum for these statically determinate trusses. It is observed from Table 4.3 that the optimal truss for Nxu = 2 has the smallest value of the objective function; therefore, it is regarded as the global optimal solution if Nxu is also considered as a variable. As demonstrated in the example above, global optimal topology can be found by searching the smallest value of C(X, A) of the optimal solutions of Problem (4.5) corresponding to all the possible sets of Nxu and Nyu . Since this process is not practically acceptable for large-scale trusses, the computational cost may be reduced if the optimal solution with smaller numbers of Nxu and Nyu can be found from the ground structure with larger numbers of Nxu and Nyu . However, there exist serious diﬃculties due to discreteness in the design variable and the assumption of uniform cross-sectional area, as shown below. 4.3.1.1

Singularity and discontinuity of optimal solutions

For a cantilever-type truss with Nyu = 1, the vertical displacement Ua of the loaded node ‘a’ and the total cost C are computed for the trusses with

Conﬁguration Optimization of Trusses

169

W 1

b

2

5

6 4

3 X0 y

c

d

9

10 8

7

e

H

12 11

a x

X1

X2

X3 Pa

FIGURE 4.6:

A plane cantilever-type truss (Nxu = 3, Nyu = 1).

TABLE 4.3:

Optimal cross-sectional area, nodal location, and total cost for Nxu = 1, 2, 3, and 4. 1 2 3 4 Nxu A 913.35 602.71 670.95 611.32 X0 0.0 0.0 0.0 0.0 X1 4000.0 2000.0 1333.3 1000.0 X2 — 4000.0 2666.7 2000.0 X3 — — 4000.0 3000.0 X4 — — — 4000.0 C 13.218 10.642 12.967 16.736

Nxu = 2 and 3 to investigate diﬃculties in ﬁnding optimal topologies by removing unnecessary nodes and members. The cross-sectional areas are 1.0 for all members of both trusses, and the nodal cost is not considered. Other parameters are the same as those in Example 4.3. The total cost for the truss with (Nxu , X1 ) = (2, 2000.0) is 1.7657, whereas that for (Nxu , X1 , X2 ) = (3, 2000.0, 2001.0) is 2.1657. The diﬀerence observed here is due to the fact that members 4, 6, and 8 in Fig. 4.6 for Nxu = 3 almost overlap with each other and the cross-sectional areas of the three members should be divided by 3 to be equivalent to the case of Nxu = 2. Next, the diﬃculty in optimizing pin-jointed trusses without rotational stiﬀness at the joints is discussed. The displacements Ua for (Nxu , X1 ) = (2, 2000.0) and (Nxu , X1 , X2 ) = (3, 2000.0, 2000.1) are 6.0271 and 6.9793, respectively, which are signiﬁcantly diﬀerent. Fig. 4.7 shows the deformation of the truss with (Nxu , X2 , X3 ) = (3, 2000.0, 2000.1). As is seen, there exists discontinuity in displacement in the y-direction between X = X1 and X2 , because the stiﬀness of member 6 in Fig. 4.6 is not inﬁnite and the thin rectangular unit ‘bdec’ deforms into a parallelogram. To prevent this discontinuity,

170

Optimization of Finite Dimensional Structures

FIGURE 4.7: Deformation of a plane truss with Nxu = 3, X1 = 2000.0, and X2 = 2000.1; solid line: deformed shape, dashed line: undeformed shape. an additional constraint may be given such that the relative displacement between nodes ‘b’ and ‘c’ as well as ‘d’ and ‘e’ should be suﬃciently small if the distance between X1 and X2 almost vanishes in the process of optimization. The diﬃculties of conﬁguration optimization of a regular truss with uniform cross-sectional area are summarized as follows: 1. The existence of extremely short members or elimination of some members leads to singularity of the stiﬀness matrix. 2. A member that has a larger cross-sectional area is generated if coalescent members are simply combined into a single member. In contrast, the global stiﬀness changes discontinuously if some among the coalescent members, leaving only one member, are removed. 3. Additional constraints are needed to restrict the relative displacement of the closely spaced nodes. 4. No node or member can be removed completely during the optimization process, because a node or a member removed once may turn out to be necessary in the ﬁnal optimal solution. 4.3.1.2

Optimization algorithm

The third diﬃculty stated above is successfully avoided by modeling the truss as a rigidly jointed frame with a suﬃciently small radius of gyration of area, as demonstrated in Sec. 4.2. To resolve the second and fourth diﬃculties, the cross-sectional area Ai of the ith member is considered to be a continuous function of X. A sigmoid function is used to deﬁne the cross-sectional areas of the closely spaced members. For instance, A∗ (= A6 = A8 ) of the truss with Nxu = 3 is deﬁned as a function of ∆X1 = X2 − X1 as ( ) ∆X1 A∗ = A tanh (4.6) sX ∗ where X ∗ is a parameter for deﬁning the nondimensional value of the distance, and s is the shape parameter of the sigmoid function; a smaller value of s leads to a shape close to the step function. By using this function, A6 and

Conﬁguration Optimization of Trusses

171

7 6

3 (1)

(2)

Y i +1

5 1

2 4

(3)

(4)

Yi

Y i −1

Xi −1

FIGURE 4.8:

Xi

Xi +1

A unit of a rectangular truss.

A8 are successfully decreased to a small value as ∆X1 is decreased, while Ai for members 4, 5, and 7 remain equal to A. This way, continuous transition between the conﬁgurations corresponding to Nxu = 2 and 3 is realized, and the second and fourth diﬃculties stated above are resolved. Since the existence of a member with null cross-sectional area or with null length may lead to unfavorable singularity of the stiﬀness matrix, side constraint ∆Xi ≥ D is given with a small positive value D. Thus, the nodes do not overlap completely, and the ﬁrst diﬃculty is avoided. For a regular plane truss as shown in Fig. 4.5, the conﬁguration of the truss is deﬁned in terms of the design variables Xi and Yi , which are the coordinates of the nodes located at the intersection of the lines in the X- and Y -directions (see Tagawa and Ohsaki (1999) for a general algorithm for non-regular plane trusses). The sigmoid functions S x and S y with respect to ∆Xi and ∆Yi are deﬁned as ( ) ( ) ∆Xi ∆Yi x y S (∆Xi ) = tanh , S (∆Yi ) = tanh (4.7) sX ∗ sY ∗ where Y ∗ is a parameter for deﬁning the nondimensional value of the distance in the Y -direction. Let Bk denote the cost of the kth node. For a rectangular unit, as shown in Fig. 4.8, Bk and Ai are calculated from B1 = BS x (∆Xi−1 )S y (∆Yi ), B2 = BS x (∆Xi )S y (∆Yi ), B3 = BS x (∆Xi−1 )S y (∆Yi−1 ), B4 = BS x (∆Xi )S y (∆Yi−1 ), A1 = AS x (∆Xi−1 ), A2 = AS x (∆Xi ), A3 = AS y (∆Yi ), A4 = AS y (∆Yi−1 ), A5 = AS x (∆Xi )S y (∆Yi ), A6 = AS y (∆Yi ),

(4.8)

A7 = AS x (∆Xi ) This way, optimization is carried out for the ground structure with suﬃciently large Nxu and Nyu , and the nodes and members are removed to obtain

172

Optimization of Finite Dimensional Structures

the optimal conﬁguration with a smaller number of units. However, for a truss that consists of a large number of units, the optimization process may converge to a local optimal solution with a larger number of units than that of the global optimal solution. In this case, an additional optimization should be carried out starting with the topology of the local optimal solution and randomly assigning the nodal coordinates. Hence, the optimization algorithm is summarized as follows: Step 1: Deﬁne the dependence of Aj and Bk on ∆Xi and ∆Yi , and assign the parameters s, X ∗ , and Y ∗ . Set parameters cm , UjU , B, and D as well as the radius of gyration of area of the members. Step 2: Randomly generate the initial values of nodal coordinates, and ﬁnd the optimal solution of the frame by using a nonlinear programming algorithm. Step 3: If the initial and ﬁnal topology as the result of optimization in Step 2 are the same, terminate the process. Step 4: Update the topology by combining closely spaced nodes and members and go to Step 2.

4.3.2

Numerical examples

Optimal conﬁgurations are found for a regular plane truss, as shown in Fig. 4.5, where E = 205.8, Pa = 98.0, W = 3000.0, H = 1000.0, X ∗ = Y ∗ = 500.0, D = 1.0, cm = 1.0 × 10−8 , and UjU = 10.0 for all displacement components. The frame consists of a sandwich section with the ﬂange distance h = 100.0. The cost of the supports is not included because the number of supports is ﬁxed during optimization and the optimal topology does not depend on the cost of supports. Note that a local optimal topology with large numbers of nodes and members is obtained if a small value is assigned for the parameter s in (4.6). In contrast, if s has a large value, the distribution of Ai and Bk are far from uniform. Therefore s is reduced linearly from 0.6 in each optimization process in Step 3. Consider a case with B = 1.0. Optimization results starting with the frame with (Nxu , Nyu ) = (6, 3) are listed in the second column in Table 4.4. As is seen, the side constraint X1 − X0 ≥ D (= 1.0) is active and the nodes at X = X1 should be removed. Similarly, the constraints X2 − X1 ≥ D, X3 − X2 ≥ D, X4 − X3 ≥ D, X6 − X5 ≥ D, Y2 − Y1 ≥ D, and Y3 − Y2 ≥ D are satisﬁed with equality, and the nodes at X = X2 , X3 , X4 , X5 , X6 and Y = Y2 , Y3 are to be removed. Therefore, a frame with (Nxu , Nyu ) = (1, 1) has been found, and obviously no further optimization is needed. Note that the total cost C in Table 4.4 includes those of intermediate small values of Ai and Bk . The optimal cross-sectional area was found for the pin-jointed truss with (Nxu , Nyu ) = (1, 1) for the purpose of veriﬁcation. The values of A and C at

Conﬁguration Optimization of Trusses

173

TABLE 4.4:

Optimal cross-sectional area, nodal location, and total cost of plane rectangular frames. (Nxu , Nyu ) (6,3) (6,3) (2,1) B 1.0 0.1 0.1 A 1352.4 889.60 893.77 X0 0.0 0.0 0.0 X1 1.0 1498.1 1500.0 X2 2.0 1499.1 3000.0 X3 3.0 1501.0 — X4 4.0 1501.1 — X5 2999.0 2999.0 — X6 3000.0 3000.0 — Y0 0.0 0.0 0.0 Y1 1498.0 1498.0 1500.0 Y2 1499.0 1499.0 — Y3 1500.0 1500.0 — C 3.7881 1.7346 1.7177

optimum are 1370.0 and 3.6925, respectively. The optimal pin-jointed truss has a little larger cross-sectional area and smaller total cost than the optimal frame, because there exist unnecessary members with small cross-sectional areas in the optimal frame. Next, optimization is carried out for a smaller nodal cost B = 0.1. Optimization results are as listed in the third column of Table 4.4. In this case, the side constraints X2 − X1 ≥ D, etc., are satisﬁed with equality, and the optimal topology of (Nxu , Nyu ) = (2, 1) has been found as a result of the ﬁrst optimization starting with the initial topology of (Nxu , Nyu ) = (6, 3). Because the global optimal truss might have a smaller number of units, another optimization was carried out starting with (Nxu , Nyu ) = (2, 1). The results are listed in the last column of Table 4.4, which veriﬁes that the optimal solution has (Nxu , Nyu ) = (2, 1). For the pin-jointed truss with (Nxu , Nyu ) = (2, 1), the optimal values of A, X1 , and C are 904.0, 1500.0, and 1.7327, respectively. Since a suﬃciently small value has been given for the radius of gyration of area of the frame model, the diﬀerences in the optimal values between the truss and the frame are moderately small. Furthermore, we can see from the examples of B = 1.0 and 0.1 that a larger value of nodal cost leads to an optimal topology with a smaller number of nodes.

174

4.4 4.4.1

Optimization of Finite Dimensional Structures

Generation of a link mechanism Introduction

So far we have considered linear inﬁnitesimal responses of trusses, where the strain-displacement relation is linear and the responses are obtained by solving the stiﬀness equation only once. Recently, extensive studies have been carried out for structural optimization under stability constraints considering geometrical nonlinearity (Ohsaki and Nakamura 1994; Ohsaki 2005a). The theoretical background and various methodologies of this ﬁeld are summarized in the book by Ohsaki and Ikeda (2007). Ohsaki and Nishiwaki (2004, 2005) developed an optimization method for generating a multistable compliant bar-joint system, which produces the speciﬁed large deformation utilizing snapthrough behavior. This method was extended to utilize enumeration of statically determinate trusses (Ohsaki, Katoh, Kinosh*ta, Tanigawa, Avis, and Streinu 2009). Sekimoto and Noguchi (2001) presented an optimization approach to trace the desired nodal path for ﬁnite-element models. Saxena (2005) developed a path generation method of a compliant mechanism using a genetic algorithm. Jutte and Kota (2008) generated an arbitrary load-displacement path using spring elements. Eriksson (2008) presented an approach considering dynamic eﬀect. In this section, the optimization method is summarized for generating a link mechanism that has the speciﬁed path of the output node due to forced displacement at the input node (Ohsaki and Nishiwaki 2007a, 2007b). Many approaches have been presented for the design of link mechanisms, mainly based on trial-and-error modiﬁcation of the design variables, including nodal locations (geometry) and member locations (topology). Kawamoto (2005) presented an approach using the truss model (Avil´es, Ajuria, Vallejo, and Hern´ andez 1997), where the nodal locations are modiﬁed to realize the desired mechanism. Kawamoto, Bendsøe, and Sigmund (2004b) utilized graph enumeration (Tuttle, Peterson, and Titus 1989), where the numbers of nodes and members should be assigned a priori (Kawamoto, Bendsøe, and Sigmund 2004a). Kim, Jang, Park, Hyun, and Nam (2005) used rigid bodies connected by springs, and the optimal topology with ﬁxed geometry is obtained by removing the unnecessary springs.

4.4.2

Mechanical model of a link mechanism

Consider a link mechanism, as shown in Fig. 4.9. A forced displacement Ua is given at the input node ‘a’, and the ﬁnal state is deﬁned such that the displacement Ub of the output node ‘b’ reaches the speciﬁed value U b : Ub = U b

(4.9)

Conﬁguration Optimization of Trusses

Ub

175

Pa a

b

Ua

FIGURE 4.9: A link mechanism; solid line: undeformed shape, dotted line: deformed shape. The solid and dotted lines in Fig. 4.9 are the conﬁgurations before and after deformation, respectively. Because the mechanism undergoes a large deformation, it is modeled as a bar-joint structure considering large nodal displacements. Consider ﬁrst a pin-jointed truss allowing elastic axial deformation of members. Let P = (P1 , . . . , Pn )⊤ and N = (N1 , . . . , Nm )⊤ denote the vectors of external loads and axial forces, where n is the number of degrees of freedom and m is the number of members. The equilibrium equation is written as DN = P

(4.10)

where the n × m matrix D is called the equilibrium matrix. The truss is characterized by the rank r of D as follows (Pellegrino and Calladine 1986): • If r = m, then the truss is statically determinate, and the vector of axial forces N for a given P is found by solving the equilibrium equation (4.10). • If r < m, then the truss is statically indeterminate, and there exist q = m − r independent self-equilibrium modes of axial forces; hence, q is called statical indeterminacy. • If r < n, then the truss is kinematically indeterminate (unstable), and there exist h = n − r independent mechanisms; hence, h is called kinematical indeterminacy. Therefore, the link mechanism that has one independent mode of deformation is classiﬁed as a truss with h = 1. For example, the mechanism in Fig. 4.9 has four members and ﬁve degrees of freedom. Hence, the kinematical indeterminacy is 5 − 4 = 1, if the equilibrium matrix is full rank. It should be noted that the mechanisms are classiﬁed into inﬁnitesimal mechanism and ﬁnite mechanism (Calladine and Pellegrino 1991; Salerno 1992; Vassart, Laporte, and Motro 2000; Garcea, Formica, and Casciaro 2005). The bar-joint structure in Fig. 4.10(a) has an inﬁnitesimal mechanism, because it is stable in the range of ﬁnite deformation that is illustrated by the dotted lines. In contrast, the structure in Fig. 4.10(b) has a ﬁnite

176

Optimization of Finite Dimensional Structures (1) 1 2 (a) Infinitesimal

FIGURE 4.10:

(2)

(1) 1 2 (b) Finite

Classiﬁcation of mechanisms; (a) inﬁnitesimal, (b) ﬁnite.

mechanism, where the member rotates freely around the support without external load. The equilibrium equations for the mechanisms in Figs. 4.10(a) and (b) are written, respectively, as follows, using the axial forces N1 and N2 as well as the nodal loads P1 and P2 : ( )( ) ( ) ( )( ) ( ) 1 −1 N1 P1 1 0 N1 P1 = , = (4.11) 0 0 N2 P2 0 0 N2 P2 As is seen, the rank of the equilibrium matrix is 1, and the kinematical indeterminacy is 1 for both cases. Therefore, only the inﬁnitesimal mechanism can be detected using the rank of the equilibrium matrix (Salerno 1992). Our purpose here is to ﬁnd the link mechanism that has the speciﬁed path of the output node under forced displacement at the input node. The approaches to this problem are classiﬁed into two categories based on multibody dynamics and ﬁnite element analysis, respectively. In the ﬁrst approach, the members are assumed to be rigid, and the path of the output node due to the forced displacement at the input node is traced using the Lagrange equations for multibody dynamics so that the length of each member is constant during deformation. The optimal shapes that minimize the error of the path of the output node from the target path can be found by considering the initial nodal locations as design variables. However, in this approach, the numbers of members and nodes should be appropriately selected and ﬁxed during optimization so that the kinematical indeterminacy is equal to 1 (Kawamoto, Bendsøe, and Sigmund 2004a). Furthermore, many geometrical constraints are needed to keep the member length constant during deformation. Therefore, this approach is not suitable for optimization that demands many analysis steps; hence, the second approach is used here (Avil´es, Ajuria, Vallejo, and Hern´andez 1997); i.e., a conventional ﬁnite element analysis of truss elements is utilized. Let Li and Ai denote the length and cross-sectional area of member i, respectively. Deﬁne C as an m×m diagonal matrix of which the (i, i)-component is equal to Ai E/Li , where E is the elastic modulus. The n×n stiﬀness matrix K is decomposed as K = DCD⊤ (4.12) The rank of K is equal to the rank r of D, because C is full-rank. Since elastic deformation is allowed for the truss at the intermediate state of optimization, the load-displacement relation is traced by a conventional

Conﬁguration Optimization of Trusses

177

displacement control or arc-length method of large-deformation analysis of trusses (Crisﬁeld 1991). A mechanism is generated through optimization under constraint such that the maximum load vanishes throughout the deformation process. For a kinematically indeterminate link mechanism with h = 1, there exists a mode (mechanism) Φ that satisﬁes KΦ = 0 corresponding to the vanishing eigenvalue of the stiﬀness matrix K. Let U denote the nodal displacement vector. An unstable structure with h = 1 can be stabilized by constraining a component of U corresponding to a nonzero component of Φ. Suppose the truss is subjected to a set of proportional loads deﬁned by P = ΛP0 with the load factor Λ and the constant vector P0 of load pattern. Then, the path of the equilibrium states in the Λ − U space, called the equilibrium path, can be traced using a displacement component as a control parameter. This way, the natural coordinate can be successfully used for large-deformation analysis (Thompson and Sung 1986; Lio, Cossalter, and Lot 2000). Note that the rank of the equilibrium matrix should be checked at the deformed conﬁguration to ensure that the truss behaves as a mechanism throughout the path. Example 4.4 Consider a two-bar link mechanism, as shown in Fig. 4.11, which is assumed to represent a snapshot of a large-deformation process. The member numbers and the numbers of displacement components are deﬁned in Fig. 4.11. The matrices D, C, and K at this deformed state are written as ( ) 2 0 −1 −1 1 1 AE AE 1 0 0 2 −1 (4.13) 1 1 , C = D= √ , K= L 0 1 2L 2 0 −1 −1 −1 1 where L is the length of the members. As is seen, the ranks of D, C, and K are 2. Suppose a proportional load is given as P = (0, 0, Λ)⊤ with the load factor Λ. The increments from the state in Fig. 4.11 are indicated by ∆. The incremental equation to be solved in the process of path-tracing analysis is given as follows, if U3 is chosen as the path parameter: 1 0 0 1 ∆U1 AE AE 1 ∆U3 0 1 0 ∆U2 = (4.14) L 2L −1 −1/2 −1/2 1 ∆Λ Because the matrix in the left-hand side has full-rank, the increments ∆U1 , ∆U2 , and ∆Λ can be successfully found for the speciﬁed increment ∆U3 as follows: 1 1 ∆U1 = ∆U3 , ∆U2 = ∆U3 , ∆Λ = 0 (4.15) 2 2 which veriﬁes instability of the mechanism.

178

Optimization of Finite Dimensional Structures 1 2

(1)

(2)

π/4

π/4 3

FIGURE 4.11:

A two-bar mechanism under deformation. Pa Initial

Intermediate Optimal Ua

FIGURE 4.12: Relation between input force Pa and input displacement Ua of initial, intermediate, and optimal solutions.

4.4.3

Problem formulation

We generate link mechanisms by removing unnecessary members as a result of optimization from a highly connected ground structure consisting of elastic truss elements. The truss is stable in the early stage of optimization, if the maximum value Pamax of the input load Pa applied in the speciﬁed direction at node ‘a’ is not zero before reaching the ﬁnal state. Therefore, in order to naturally generate an unstable mechanism, a constraint is given such that Pamax vanishes. Fig. 4.12 shows the load-displacement relation at the input node from the initial undeformed state to the ﬁnal state for the initial, intermediate, and optimal solutions, which illustrates that the maximum load is reduced to 0 through optimization. Since the maximum load is always nonnegative, as seen in Fig. 4.12, the constraint on the maximum load is given with inequality as Pamax ≤ 0

(4.16)

which is satisﬁed with equality at the optimal solution with a small tolerance ε as Pamax ≤ ε. The design variables are the vectors of cross-sectional areas A and the nodal coordinates X. In order to obtain a mechanism with a small number of members, the total structural volume V (A, X) is minimized as the objective function. If we assign the constraint (4.16) only, then all the cross-sectional areas vanish as a result of minimizing the total structural volume. Therefore,

Conﬁguration Optimization of Trusses

179

the stiﬀness constraints are given in terms of the nodal displacements as f0

f0

f0 f0 Ubx ≤ U bx , Uby ≤ U by

(4.17)

f0 f0 are the displacements of the output node ‘b’ in the xand Uby where Ubx and y-directions against the unit loads in the x- and y-directions at node ‘b’, f0 f0 respectively, after ﬁxing the input degree of freedom, and U bx and U by are their speciﬁed upper bounds. Note that the tangent stiﬀness at the deformed f0 f0 conﬁguration is used for evaluating Ubx and Uby (Ohsaki and Nishiwaki 2005). The optimization problem for ﬁnding the link mechanism is formulated as

Minimize

V (A, X)

subject to

Pamax (A, X)

(4.18a) ≤0

f0 Ubx (A, X) ≤ f0 Uby (A, X) L

≤

f0 U bx f0 U by

A ≤A

X ≤X≤X L

(4.18b) (4.18c) (4.18d) (4.18e)

U

(4.18f)

where XU and XL are the upper and lower bounds for X, respectively, and AL = (AL1 , . . . , ALm )⊤ deﬁnes the small lower bound for A. Note that the member with Ai = ALi at the optimal solution is to be removed. The optimization algorithm is given as follows: Step 1: Assign initial values for A and X. Step 2: Trace the equilibrium path using the input displacement at node ‘a’ as the control parameter until reaching the ﬁnal state satisfying condition (4.9). Step 3: Evaluate the objective and constraint functions and their sensitivity coeﬃcients. Step 4: Update the variables according to the optimization algorithm. Step 5: Go to Step 2 if the convergence criteria are not satisﬁed. In Step 2, the optimization process is terminated if the output displacement is not in the speciﬁed direction, because, in this case, the desired mechanism is not likely to be obtained. Furthermore, the optimization problem is highly nonlinear, and many local optimal solutions are expected to exist. Therefore, the initial values of A and X are randomly generated, and path tracing is continued only for the case where the output displacement Ub is in the speciﬁed direction. The optimal solution of the ﬁrst-stage problem (4.18) is unstable and sateb isﬁes the requirements for a link mechanism. However, the displacement U

180

Optimization of Finite Dimensional Structures 1

Input

Output

a

a Output b

2

1 2

Input d

c

(a) Case 1 FIGURE 4.13:

(b) Case 2

Examples of unnecessary members.

of the output node in the direction perpendicular to the speciﬁed direction may be large, because no constraint has been given on the direction of the output displacement. Therefore, in the second-stage problem, the maximum e max of U eb is minimized by considering X as design variables, absolute value U b while A is ﬁxed: Minimize

ebmax (X) U

subject to X ≤ X ≤ X L

(4.19a) U

(4.19b)

Hence, the link mechanisms with the desired property can be obtained as the solution of the two-stage problem. Note again that the unnecessary members are removed in the ﬁrst-stage problem for minimizing the total structural volume. However, because the problem is highly nonlinear, strict convergence to the global optimal solution may not be expected. Even when the global optimum is not found, a feasible solution is obtained in the ﬁrst stage, and we can proceed to the second stage. Since the topology is ﬁxed at the second stage, unnecessary members cannot be removed at this stage. However, an unnecessary member that can be removed without any eﬀect on the load-displacement relation may be detected manually utilizing the following criteria: 1. If no stress exists in a member against the input force after constraining the output displacement, then the member is considered to be unnecessary. For example, if only two members, which are not colinear, are connected to a node except the input node, output node, and supports, these members are unnecessary, e.g., node ‘a’ in Fig. 4.13(a). 2. If the global property does not change after removing a member in a statically indeterminate substructure, then the member is unnecessary. For example, the quadrilateral substructure consisting of nodes ‘a’, ‘b’, ‘c’, and ‘d’ is statically indeterminate, and members 1 or 2 can be removed without loss of stability of the substructure. Note that some members can also be added to replace the existing members to simplify the topology. The performance of the structure after removing the members is to be conﬁrmed by path-tracing analysis.

Conﬁguration Optimization of Trusses

181

y 1

H 3

H 4

2 W

FIGURE 4.14:

4.4.4

W

x

W

A 3 × 2 plane grid truss.

Numerical examples

Link mechanisms are found from the ground structure shown in Fig. 4.14, where W = H = 200 mm. The truss is pin-supported at node 1 and ﬁxed in the x-direction at node 2. The ground structure has a 3 × 2 grid, and the intersecting diagonals are not connected with each other. We obtain the mechanism so that node 4 moves 200 mm in the y-direction as the result of anticlockwise rotation of node 3 around support 1. For this purpose, a large value is given for the lower bound ALi of the cross-sectional area of the thick member in Fig. 4.14 connecting support 1 and node 3. Because the optimal solution is unstable, the value of Ai does not have any eﬀect on the performance of the link mechanism obtained after optimization of cross-sectional areas and nodal locations. The anticlockwise rotation is realized by the forced y-directional displacement at node 3. Hence, the ﬁnal state is deﬁned as the deformation where the output displacement Ub in the y-direction of node 4 reaches 200 mm as the result of y-directional input displacement Ua at node 3. The upper-bound displacement against unit loads f0 f0 (1.0 N) in the ﬁrst stage is given as U bx = U by = 50 mm. The displacement in the x-direction of node 4 is minimized in the second stage optimization problem (4.19). The cross-sectional areas of members, except the thick member connecting support 1 and node 3, are considered as independent variables, for which the lower bounds are equal to 0.01 mm2 . The x-coordinate of node 3, the (x, y)coordinates of node 4, and the coordinates in the constrained directions of the supports are ﬁxed, and the remaining coordinates are the independent design variables. In order to prevent distortion of the topology, the bounds in the side constraints (4.18f) for the nodal coordinates are given as the square

182

Optimization of Finite Dimensional Structures

(a) Initial state

(b) Deformed state

Horizontal displacement

2 1 0 -1 -2 0

50

100

150

200

Vertical displacement

(c) Relation between vertical and horizontal displacements (mm) of node 4 FIGURE 4.15: Optimal link mechanism before removing unnecessary members.

(a) Initial state FIGURE 4.16: bers.

(b) Deformed state

Optimal link mechanism after removing unnecessary mem-

region bounded by ±60 mm from the original locations deﬁned in Fig. 4.14. The engineering strain of a member is computed exactly from the deformed 2 and undeformed lengths of the member. The elastic modulus E is 2.0 N/mm . Because the ﬁnal mechanism is unstable, the value of E does not have any eﬀect on the optimal solution. The y-directional displacement of node 3 is

Conﬁguration Optimization of Trusses

183

chosen as the parameter for path-tracing analysis. The performance of the mechanism is conﬁrmed by ADAMS 2005 (MSC Software 2005), which is a general purpose software package for multibody dynamics. The sequential quadratic programming program in IDESIGN Ver. 3.5 (Arora and Tseng 1987) is used for optimization, where the tolerance for the constraints is ε = 1.0 × 10−4 . The sensitivity coeﬃcients are evaluated by a ﬁnite diﬀerence approach. In order to obtain various mechanisms, initial solutions are generated randomly. Since the ratios rather than the absolute values are important for the cross-sectional areas, the initial values are deﬁned so as to distribute uniformly between 0 and an appropriate positive value A0 as Ai = A0 Ri , where 0 ≤ Ri < 1 is a uniform random number, and A0 = 10 mm2 in the following examples. The initial value for the nodal coordinate xi (mm) of the ith node is given as xi = x0i + 120.0(Ri − 0.5) (4.20) where x0i is the coordinate in Fig. 4.14. The y-coordinates are deﬁned in a similar manner. A mechanism obtained by using the proposed two-stage optimization algorithm is shown in Fig. 4.15(a). Fig. 4.15(b) shows the deformed ﬁnal shape, where the dashed horizontal line is added to clearly indicate that the y-coordinate is ﬁxed at support 1. Fig. 4.15(c) shows the relation between y-directional and x-directional displacements of output node 4. Note that the horizontal and vertical axes of the ﬁgure correspond to vertical (y-directional) and horizontal (x-directional) displacements, respectively, and their scales are diﬀerent; therefore, node 4 moves almost straight in the y-direction. The shape after removing the unnecessary members is shown in Fig. 4.16. The relation between y-directional and x-directional displacements of node 4 is the same as Fig. 4.15(c). The dotted members have been added to obtain a simple conﬁguration. Note that the members that can be added have been assumed to be selected from the ground structure. However, a simpler conﬁguration can be obtained if members can be added between any pair of nodes. The numbers of members and degrees of freedom are 18 and 19, respectively, after removing the unnecessary members. Therefore, the kinematical indeterminacy h is equal to 1, if the equilibrium matrix is full-rank. Several mechanisms can be obtained by carrying out optimization from diﬀerent initial solutions (Ohsaki and Nishiwaki 2007a).

Chapter 5 Optimization of Building Frames

Optimum design of building frames was extensively studied in the 1960s, because analytical derivation of optimal solutions is possible for simple regular frames. Since the 1980s, with rapid development of computer technology, many methodologies and computer programs have been presented for optimizing real-world structures under practical design constraints. In this chapter, we present various optimization results for building frames. A historical review is presented in Sec. 5.1. The uniqueness of optimal solutions is investigated and approximate solutions are searched for in Sec. 5.2. Multidisciplinary optimization and parametric optimization approaches are presented in Sec. 5.3. A heuristic approach to multiobjective optimization is presented in Sec. 5.4, and ﬁnally, multiobjective seismic design optimization is discussed in Sec. 5.5.

5.1 5.1.1

Overview of optimization of building frames Introduction

Optimum plastic design of building frames is a traditional problem investigated in the 1960s, because the optimal solutions are found analytically or by solving a linear programming problem, as we have seen in Sec. 1.4 (Hemp 1973). Various methods were developed in the 1970s mainly using optimality criteria approaches; see Sec. 1.7 for elastic design problems considering stress and displacement constraints. However, practical design of building frames using sophisticated computational optimization techniques, including nonlinear programming (NLP) and heuristics, is a rather new ﬁeld of research in spite of its importance in design practice, because various constraints and costs in view of structural and nonstructural requirements should be considered. Arora, Haug, and Rim (1975) applied a state-space optimal control approach to optimization of frames. Bhatti and Pister (1981) may be cited as one of the ﬁrst papers on practically applicable optimization methods of building frames. Balling, Pister, and Ciampi (1983) presented an NLP approach under dynamic constraints, where geometrical nonlinearity is considered. Sui and Wang (1997) used geometric programming for frame optimization. Lagaros, Papadrakakis, and

185

186

Optimization of Finite Dimensional Structures

Kokossalakis (2002) presented evolutionary approaches for frame optimization. J´armai, Farkas, and Kurobane (2006) used particle swarm optimization (PSO) (Kennedy 1997) for optimization of a plane steel frame, where the practical objective function, including the costs of connection and fabrication, is minimized under code-based design constraints. PSO has also been applied to the design of trusses (Li, Huang, Liu, and Wu 2007). Saka and Erdal (2009) applied a harmony search algorithm for frame optimization based on load and resistance factor design (LRFD) speciﬁcations. Optimum design of real-world tall buildings has been studied since the 1980s. Grierson and Lee (1984) presented a two-level approach, where the problem with continuous variables is ﬁrst solved to select the initial crosssections for the second problem with discrete variables. Grierson and Chan (1993) developed an optimality criteria approach to the optimization of steel frames under lateral drift constraints, where the nodal displacements are explicitly written using the virtual force method, and the cross-sectional properties of each member are approximated as linear functions of the reciprocal of the cross-sectional area. Chan, Grierson, and Sherbourne (1995) optimized a 50-story steel frame using an optimality criteria approach to select the crosssectional properties of members from a list of commercially available sections. Haque (1996) applied the complex method (Box 1965) for frame optimization with discrete variables. Kocer and Arora (1999) optimized a transmission tower using a simulated annealing and a genetic algorithm, where the material grades as well as the cross-sectional properties are considered as discrete design variables. Liu, Burns, and Wen (2006) presented a multiobjective programming approach for minimization of the number of diﬀerent types of cross-sections. Bojczuk and Mr´oz (1998a) optimized the support locations of a continuous beam. Recently, some attempts have been made to optimize the cross-sectional shapes of beams. Pan, Ohsaki, and Tagawa (2007) and Ohsaki, Tagawa, and Pan (2009) optimized the ﬂange shape of a beam considering plastic energy dissipation under monotonic and cyclic loads. Erdal and Saka (2009) optimized the shape of the opening in the web of an I-beam using a heuristic approach called harmony search. Lagaros, Psarras, Papadrakakis, and Kokossalakis (2008) investigated the eﬀect of the web opening on the optimum design of 3D frames.

5.1.2

Problem formulation

Consider a plane frame, e.g., as shown in Fig. 5.1, consisting of rigidly jointed beam-column (frame) elements. The assumption of a rigid ﬂoor is usually used; i.e., the horizontal displacements of the nodes in the same ﬂoor have the same value. Note that rigid ﬂoor means that the ﬂoor is rigid in the in-plane direction, and does not mean that it is rigid against out-of-plane bending/shear deformation. Let n and nF denote the number of degrees of freedom and the num-

Optimization of Building Frames W

W

W/2

P6

W

W

W

187

W/2

W/2

W

W

W

W

W

W/2

4m

W/2

W

W

W

W

W

W/2

4m

W/2

W

W

W

W

W

W/2

4m

W/2

W

W

W

W

W

W/2

4m

W/2

W

W

W

W

W

W/2

4m

P5 P4 P3 P2 P1

4m

8m

8m

FIGURE 5.1:

8m

A six-story three-span frame. 2 3 a b

FIGURE 5.2:

5 1

c d

4

6

Displacement numbers of a member of a plane frame.

ber of stories, respectively. The nodal load vector is denoted by PN = (P1N , . . . , PnN )⊤ , which includes the vertical static loads and the horizontal loads P1 , . . . , PnF representing the seismic loads and the wind loads applied at the ﬂoors. The vertical loads include the self-weight of the beams and columns, the weights of the nonstructural components, and the live loads. Note that the horizontal load Pi−1 is applied at the ith ﬂoor, as shown in Fig. 5.1, and the roof is denoted as the (nF + 1)st ﬂoor, at which the load PnF is applied. The n×n stiﬀness matrix is denoted by K. Then the displacement vector U is found from the stiﬀness (equilibrium) equation KU = PN

(5.1)

The displacement vector of member i is denoted by ui = (ui1 , . . . , ui6 )⊤ , where the local displacement numbers are indicated in Fig. 5.2. The vector of stresses at the member ends ‘a’, ‘b’, ‘c’, and ‘d’ in Fig. 5.2 is denoted by σ i = (σai , σbi , σci , σdi )⊤ , which is written in terms of ui as σ i = di ui

(5.2)

where di is a 4 × 6 constant matrix. In most design codes, such as the Japanese Building Standard Law, the bound for the stress for a plane frame is given as follows as a summation of

188 stress ratios:

Optimization of Finite Dimensional Structures |σ A | |σ B | + B ≤1 σA σ

(5.3)

where σ A is the stress due to axial force, σ B is the stress due to bending moment, and σ A and σ B are their upper bounds. Note that diﬀerent formulas and bounds are assigned in design codes for tensile and compressive states distinguished by the sign of σ A . For a three-dimensional frame, the stresses due to bending moments around the two principal axes, denoted by subscripts ( · )y and ( · )z , are incorporated as |σ A | |σyB | |σzB | + B + B ≤1 σA σy σz

(5.4)

One of the important aspects of the design of building frames is that the cross-sectional properties, including the area and second moment of inertia, cannot be deﬁned as independent continuous variables; they should be selected from the catalog of available standard sections. Therefore, the optimization problem turns out to be a combinatorial problem. An example of the list of available sections is shown in Appendix A.8. However, approximate optimal solutions are usually found by considering cross-sectional parameters as continuous variables. Let A = (A1 , . . . , Am )⊤ denote the vector of cross-sectional areas of the members, where m is the number of members, including the beams and columns. In order to ﬁnd an optimal solution using a formulation of the nonlinear programming (NLP) problem, considering A as continuous design variables, it is necessary that the cross-sectional properties, e.g., second moment of inertia and the section modulus of the ith member, are given as functions of Ai as Ii (Ai ) and Zi (Ai ), respectively. Approximation is also needed for the plastic modulus Zip if inelastic responses are to be considered. These values are usually approximated as Ii (Ai ) = aAbi for which the parameters a and b are to be identiﬁed from the list of sections (see Appendix A.8 for details). Note that these parameters should be found so that they accurately approximate the series of sections that is to be used in the respective design problem. Therefore, it is not important here to show the values of a and b used in the literature. Alternatively, Zi and Ii may be linearly interpolated with respect to Ai if the list of sections is given so that Zi and Ii are increasing functions of Ai . Although, in this case, the sensitivity coeﬃcients of Zi and Ii with respect to Ai are discontinuous, optimal solutions are found successfully, as demonstrated in Example 5.1, if an NLP algorithm with line search is used. In the design process of building frames, the upper bounds are usually given for the response displacements and stresses against static loads based on the design code. Therefore, the design requirements are generally formulated as follows as inequality constraints: Hj (A) ≤ 0, (j = 1, . . . , nI )

(5.5)

Optimization of Building Frames

189

where Hj is assumed to be an implicit function of A after eliminating the state variables, and nI is the number of constraints. Hence, the optimization problem for minimizing the objective function, e.g., the total structural volume V (A), is formulated as a standard NLP problem: Minimize

V (A)

(5.6a)

subject to Hj (A) ≤ 0, (j = 1, . . . , nI ) ALi

≤ Ai ≤

AU i ,

(i = 1, . . . , m)

(5.6b) (5.6c)

where ALi and AU i are the lower and upper bounds for Ai , respectively. The deﬁnition of the objective function based on the cost of the structure is the most ambiguous aspect of structural optimization in the ﬁeld of civil engineering. Various deﬁnitions have been proposed, including structural and nonstructural costs (see Sec. 5.1.5 for details). The most realistic deﬁnition of the cost may be the life-cycle cost; however, the possible hazards and maintenance costs during the life cycle of the structure are highly unpredictable. Hence, it is very diﬃcult to present a realistic and robust deﬁnition of the lifecycle cost. Therefore, the initial cost that is represented by the volume of the structural material is used for the objective function of a simple optimization problem of a building frame. Another diﬃculty in the optimization of building frames is that various short-term loads, including seismic, wind, and snow loads, represented by equivalent static loads, should be considered in addition to long-term loads, including self-weight and live load. An optimum design should be found so that the structure remains in the elastic range for the long-term load and has enough safety with inelastic deformation against short-term loads (Grierson and Schmit 1982; Grierson and Chiu 1984). This concept was recently generalized as performance-based design, for which is the design process can be naturally formulated as an optimization problem (Foley 2002). In view of practical application, some additional requirements should be considered for the construction process (Liu, Burns, and Wen 2006). For instance, construction cost depends on the complexity of the structure such as the number of diﬀerent sections, number of parts to be assembled, and shape of the parts at the connections. Furthermore, requirements on widththickness ratios of the ﬂanges and webs of beams are given by the building codes. The strong column-weak beam criterion should usually be satisﬁed; i.e., the frame should have enough capacity of plastic energy dissipation at the plastic hinges at the beam ends before exhibiting total collapse due to the failure of columns. Example 5.1 As a small example, optimal solutions are found for the six-story three-span plane steel frame shown in Fig. 5.1 (Ohsaki 2003a). The conventional assumption of a rigid ﬂoor is used. The vertical load W = 120 kN represents the

190

Optimization of Finite Dimensional Structures

TABLE 5.1:

List of available sections for beams and columns; cross-sectional area A (×102 mm2 ), section modulus Z (×103 mm4 ), and second moment of inertia I (×106 mm6 ). Beam Column A Z A Z I I G1 61.4 610.0 106.0 C1 104.0 986.0 147.0 G2 83.4 1040.0 232.0 C2 162.0 1760.0 309.0 G3 116.0 1550.0 387.0 C3 186.0 2330.0 467.0 G4 142.0 2180.0 655.0 C4 277.0 3880.0 873.0 G5 184.0 3040.0 896.0 C5 309.0 4840.0 1210.0 G6 214.0 3560.0 1050.0 C6 599.0 8780.0 2190.0

2

self-weight and live load, and the elastic modulus is 200 kN/mm . The horizontal loads (kN) representing the seismic loads are given as (P1 , . . . , P6 ) = (60.0, 110.0, 155.0, 195.0, 230.0, 260.0). The sections of beams and columns are selected from the list in Table 5.1, where A, Z, and I are the cross-sectional area, section modulus, and second moment of inertia, respectively. The total structural volume V is to be minimized. In the following, the units of length and force are mm and kN, respectively, which are omitted for brevity. The columns and beams are modeled using Euler-Bernoulli beam elements, and each beam is divided into two elements in order to incorporate the stress constraints at the center of the beam. The upper bound 0.2 is given for the absolute value of the stresses at the edges of the two ends, denoted by ‘a’, ‘b’, ‘c’, and ‘d’ in Fig. 5.2, of the elements, including the beams and columns. In order to incorporate a realistic situation and to reduce the number of design variables, the beams and columns are classiﬁed into six groups, respectively, where the members in each group have the same section. The columns are classiﬁed into exterior and interior columns, and those in three pairs of stories (1,2), (3,4), and (5,6) have the same sections. The beams are classiﬁed into six groups similarly. We ﬁrst interpolate Z and I as piecewise linear functions of A using the values in Table 5.1. Note that the available sections have been chosen such that Z and I are increasing functions of A for both the columns and beams. The optimization software package IDESIGN Ver. 3.5 (Arora and Tseng 1987), which utilizes sequential quadratic programming, is used. The bounds for the cross-sectional areas are deﬁned by the smallest and largest values in the lists for beams and columns, respectively. The optimal solution denoted by AR with continuous variables is listed in the second column of Table 5.2, where, e.g., ‘Ext. beam 2, 3’ indicates the beams in the exterior spans in the second and third ﬂoors, and ‘Int. col. 1, 2’ indicates the interior columns in the ﬁrst and second stories, while the roof is denoted as the seventh ﬂoor. The total structural volume is also listed in Table 5.2, and the ‘Max. stress’ means the maximum absolute value of the stresses among all members. Note that

Optimization of Building Frames

TABLE 5.2:

Ext. beam 2, 3 Ext. beam 4, 5 Ext. beam 6, 7 Int. beam 2, 3 Int. beam 4, 5 Int. beam 6, 7 Ext. col. 1, 2 Ext. col. 3, 4 Ext. col. 5, 6 Int. col. 1, 2 Int. col. 3, 4 Int. col. 5, 6 Total volume (×106 ) Max. stress

191

Optimal solutions of the six-story three-span frame. Cont. Greedy from Enum. Greedy Stingy var. AR nearest near AR to AR 214.0 G6 G6 G6 G6 188.8 G6 G5 G5 G6 121.1 G3 G3 G4 G4 206.2 G6 G6 G6 G6 186.7 G5 G6 G6 G6 122.2 G5 G4 G3 G4 289.1 C4 C4 C4 C5 186.2 C3 C3 C3 C3 116.2 C1 C3 C2 C2 404.3 C6 C6 C6 C6 282.8 C4 C4 C4 C5 175.5 C3 C3 C2 C3 4.827 5.278 5.259 5.538 5.182 0.2

0.1971

0.1986

0.1949

0.1987

the stress constraint is satisﬁed with equality in at least one of the members in each group; i.e., the optimal solution is fully stressed. An approximate optimal solution from the list of available sections can be found by selecting the nearest section from AR . If the stress constraint is not satisﬁed, then the section of the group with maximum violation is consecutively increased until the stress constraints are satisﬁed in all members. The solution obtained by using this simple greedy method is listed in the third column of Table 5.2; see Appendix A.3 for details of the greedy method. The maximum stress is 0.1971; i.e., the solution is not fully stressed, because the sections can be selected only from the discrete candidates. The total structural volume is 5.278 × 106 , which is about 9.3% larger than AR . If the available sections are restricted to the two nearest sections from AR for each group, then we have 212 = 4048 candidate solutions. The optimal solution obtained by enumeration of these candidate solutions is listed in the fourth column of Table 5.2. Note that the total structural volume is almost the same as that found by the simple greedy method, although structural analysis has been carried out 4048 times. We next use the greedy method starting with the smallest section for each group. The approximate solution obtained this way is shown in the ﬁfth column of Table 5.2, which is worse than the previous two discrete solutions. In contrast, if we use the stingy method starting with the largest section for each group, the best approximate solution is obtained as shown in the sixth column of Table 5.2. Although the performances of the methods depend on the deﬁnition and parameter values of the problem, good approximate discrete

192

Optimization of Finite Dimensional Structures

solutions can be obtained using heuristic approaches.

5.1.3

Continuum approach

There are numerous papers on topology optimization of plates (sheets) discretized to ﬁnite elements subjected to in-plane loads. As a result of optimization, material layouts like frame structures, e.g., bridges and braced frames, are often obtained. Therefore, some attempts have been made to generate optimal layout of beams, columns, and braces of building frames utilizing continuum shape and topology optimization. Yang, Xie, Steven, and Querin (1999b) used the bi-directional evolutionary structural optimization (BESO) method (Yang, Xie, Steven, and Querin 1999a) for frequency optimization of frames, which was extended from the original evolutionary structural optimization (ESO) method allowing addition of the elements. Liu and Qiao (2009) optimized the layout of frames and bridge-type structures with diﬀerent stiﬀnesses in tension and compression that are approximated by a smooth function. Mijar, Swan, Arora, and Kosaka (1998) presented a continuum optimization approach for generating braced frames utilizing the Voigt-Reuss mixing rule of constitutive relation. Rahmatalla and Swan (2003) developed a continuum-based conceptual design approach for bridges and transmission towers. However, we do not discuss continuum approaches in this book, because there are many books on continuum shape/topology optimization (Haug and Cea 1981; Bendsøe and Sigmund 2003).

5.1.4

Semi-rigid connections and braces

Semi-rigid connections are widely used in American and European countries to concentrate the resistance to horizontal loads at the limited number of spans consisting of moment-resisting (rigidly jointed) frames. This way, the total cost of a frame is reduced, because the manufacturing and construction of rigid joints consisting of welding or high-tension bolts are more expensive than those of semi-rigid connections. In contrast, semi-rigid connections are not very popular in Asian communities, especially Japan, where columns with a box section are usually used with diaphragms welded to the ﬂanges of the beams. However, rigorously speaking, most of the connection types are semirigid, and there is no strictly pin-jointed or rigidly jointed connection. Various results of optimization of frames with semi-rigid connections are summarized by Xu (2002). Machaly (1986) optimized the cross-sectional dimensions of the members of an elastic frame with semi-rigid connections. Kameshki and Saka (2001b) and Hayalioglu and Degertekin (2005) used genetic algorithms for optimizing frames with semi-rigid connections, where the types of connections are ﬁxed. In contrast, Kameshki and Saka (2003) also optimized the types of the semi-rigid connections. The Frye–Morris polynomial model (Frye and Morris 1975) is often used

Optimization of Building Frames 2

5 1

3

FIGURE 5.3:

193

4 6 node 2

node 1

A beam-column element with semi-rigid connections.

for the moment-rotation relation of a semi-rigid connection. Machaly (1986) showed that the use of semi-rigid connections instead of rigid connections leads to a reduction of the total weight of a small-scale frame. Xu and Grierson (1993) used a continuous model for the stiﬀness of the connections. Doˇgan and Saka (2009) used a particle swarm optimization method for optimizing a frame under an AISC-LRFD speciﬁcation (AISC 1999). Hagish*ta and Ohsaki (2008a) developed a heuristic approach based on scatter search for optimization of braced frames with semi-rigid connections considering inelastic static responses. Kameshki and Saka (2001a) optimized the locations of braces using a genetic algorithm. Kameshki and Saka (2001b) optimized the types of braces of frames with semi-rigid connections. Consider a beam-column element connected to semi-rigid connections (nodes 1 and 2) with elastoplastic springs, as shown in Fig. 5.3. The elastic modulus, second moment of inertia, and length of the element are denoted by E, I, and L, respectively. Let Ri denote the rotational stiﬀness of of the spring at node i, and deﬁne parameters ra , rb , rc , a, b, and c as ( ) 1 12EI 4 + R∗ LRj ( ) 1 12EI rb = ∗ 4 + R LRi 2 rc = ∗ R a = ra + 2rb + rc b = ra + rb c = rb + rc ra =

(5.7a) (5.7b) (5.7c) (5.7d) (5.7e) (5.7f)

where R∗ =

( 1+

4EI LRi

)( 1+

4EI LRj

)

( −

EI L

)2 (

4 Ri Rj

) (5.8)

The local displacement numbers are deﬁned in Fig. 5.3. Let A denote the cross-sectional area of the member. Then the 6 × 6 elastic stiﬀness matrix KE

194

Optimization of Finite Dimensional Structures

of a member of a plane frame is deﬁned as (Chen 1991) EA L KE =

EI a 3 L

EI b 2 L EI ra L

−

EA L 0 0

EA L sym.

0

EI EI −a 3 c 2 L L EI EI −b 2 rc L L 0 0 EI EI a 3 −c 2 L L EI rb L

(5.9)

For performance evaluation of the frames with semi-rigid connections, it is very important to consider inelastic responses, because the large rotational deformation is concentrated at the connections. For this purpose, the reﬁned plastic hinge method (Liew, White, Chen, and Toma 1993a, 1993b), which is categorized as advanced analysis, is often used. This method incorporates geometrical and material nonlinearities, as follows, which are equivalent to the load and resistance factor design (LRFD) speciﬁcation (AISC 1999; Chen and Kim 1997). The geometrically nonlinear deformation of a two-dimensional beam-column element is modeled using corotational formulation (Crisﬁeld 1991) so that the local coordinates are rotated in accordance with the global deformation and one of the local axes is always in the direction of the member axis. Hence, the 6 × 6 stiﬀness matrix of a beam-column element in global coordinates is reduced to a 3 × 3 matrix with respect to local coordinates. Let δMi and δθi denote the incremental bending moment and rotation, respectively, of node i of the element. The increments of axial force and elongation are denoted by δP and δe, respectively. Then, the incremental form of the equilibrium (stiﬀness) equation is written as δM1 S1 S2 0 δθ1 EI δM2 = S2 S1 0 δθ2 L δP 0 0 A/I δe

(5.10)

where S1 and S2 are the stability functions to incorporate the secondary force and moment due to P − δ and P − ∆ eﬀects (Chen and Kim 1997). In order to incorporate the gradual yielding due to nonuniform distribution of the bending moment, the perfectly rigid-plastic model can be modiﬁed to an elastoplastic model in a manner similar to the LRFD formula. There have been many models proposed for the relation between the moment M and rotation angle θ of semi-rigid connections. For example, Kishi

Optimization of Building Frames

195

TABLE 5.3:

Three parameters for ﬁve types of semi-rigid connections. Connection type Mu Rinit Shape (kNm) (kNm/rad) parameter s Single web angle 18.4 164.1 1.8 Double web angle 36.8 3282.2 2.7 Top and seat angle 54.0 23,409.4 0.8 Top and seat angle with 93.9 26,786.8 1.2 double web angles T-stub 150.0 60,000.0 2.0

and Chen (1990) presented the following simple three-parameter model: θ θe Mu M , θe = , θ0 = = 1 s e Mu θ R s 0 init (1 + θ )

(5.11)

where the parameters Mu , Rinit , and s are the ultimate bending moment, initial rotational stiﬀness, and shape parameter, respectively. Semi-rigid connections are classiﬁed into ﬁve types: (1) single web angle, (2) double web angle, (3) top and seat angle, (4) top and seat angle with double web angles, and (5) T-stub. Types (1)–(4) use angle members at diﬀerent locations for connecting the beam and the column. Simple formulas have been developed for the three parameters of these types, as shown in Table 5.3. Frye and Morris (1975) presented the following odd-power polynomial model for the connections (a) end plate without column stiﬀness, (b) top angle seat with double web angle, and (c) top angle seat without double web angle: θ = C1 (KM ) + C2 (KM )3 + C3 (KM )5

(5.12)

where Ci (i = 1, 2, 3) are the curve-ﬁtting parameters for each type of connection, and K is a parameter deﬁned in view of the geometry of the connection. Hagish*ta and Ohsaki (2008a) optimized a frame with semi-rigid connections and braces, as shown in Fig. 5.4, using a heuristic approach called scatter search (Laguna and Marti 2003), where the units of load and length are kN and mm, respectively. The types of semi-rigid connections are selected from the ﬁve models in Table 5.3. The locations, sections, and types of the braces, including X-brace, K-brace, and V-brace, are also optimized. Note that any type of brace can exist at each span of the story. The frame is subjected to vertical loads (long-term loads) and two types of horizontal loads (seismic loads), where the horizontal seismic loads of Levels 1 and 2 are shown, respectively, without and with parentheses in Fig. 5.4. The vertically aligned columns have the same section, and the standard assumption of a rigid ﬂoor is used. The sections of beams, columns, and braces are selected from the list of the available standard sections. Constraints are given for the interstory

Optimization of Finite Dimensional Structures

139.40 (418.25)

22.5

45.0 56.25

67.5

56.25 45.0

56.25 67.5 45.0 112.68 22.5 (338.08) 67.5 45.0 56.25 96.42 22.5

56.25 45.0

67.5

56.2545.0

45.0 56.25 83.46 22.5 (250.43)

56.25 45.0

4500

(289.31)

9000

6000

3 @ 3500

196

9000

6000

6000

FIGURE 5.4: Geometry and loading conditions of a braced frame with semi-rigid connections. 5

5 3

3 2

4

4 2

2

1

1 1

1

1

1 2

1. Single web angle

2

2 3

35

5

2

2 4

4

1

1 1

1

1

1

5

5

1 3

1 3 2

2. Double web angle

4. Top and seat angle with double web angle

FIGURE 5.5:

2

3. Top and seat angle 5. T-stub

Optimal types of braces and semi-rigid connections.

drift angle under the seismic loads of Level 1 and the maximum load carrying capacity (ultimate plastic collapse load) under Level 2; see Hagish*ta and Ohsaki (2008a) for details. The cost of each semi-rigid connection is added to the total structural volume to formulate the objective function. An example of an optimal solution is shown in Fig. 5.5, where the width of each member is proportional to the height of the cross-section, and the types of semi-rigid connections are also indicated. As is seen, the beams in the short span have large sections. The braces are located so as to form a hat truss in the fourth story and to transmit the horizontal loads diagonally to the supports between the ﬁrst and third stories.

Optimization of Building Frames

5.1.5

197

Formulation of cost function

One of the most diﬃcult aspects in the application of optimization approaches to practical design of building frames is that it is very diﬃcult to formulate the objective (cost) function incorporating the various structural and nonstructural costs. As discussed in Sec. 1.1, we can obtain feasible designs through optimization even if the deﬁnition of the cost function is unrealistic. However, it is desired that good approximate optimal solutions are found with a realistic deﬁnition of the cost function. Pavlovˇciˇc, Krajnc, and Beg (2004) investigated in detail the initial cost of a steel frame and classiﬁed it into the costs for steel material, welding manufacturing, welding material, assembly and trach welding, cutting, painting, surface preparation, ﬂange aligning, joint, hole forming, bolting material, transportation, and erection. However, it is not possible to precisely deﬁne these costs as functions of the design variables. One of the criticisms of the optimization of building frames is that it considers only the initial cost, or only the material cost, and does not include the cost throughout the service life of the building. Recently, some attempts have been made to optimize the frame considering the life-cycle cost. Sarma and Adeli (2002) presented a detailed classiﬁcation of the cost and deﬁned the life-cycle cost as the sum of the costs of initial construction, maintenance, inspection, repair, operation, failure, and dismantling. Doi, Yoshimura, Nishiwaki, and Izui (2009) classiﬁed industrial products into short/long functional longevity and short/long usage period, where building frames are classiﬁed into the product with long usage period. They considered the design for ‘3R’, i.e., reduce, reuse, and recycling, and formulated the total cost as the sum of manufacturing cost, recycling cost, and reuse cost. Wen and Kang (2001) optimized the frame considering the life-cycle cost under multiple hazards. Liu, Wen, and Burns (2004) presented a multiobjective optimization approach considering the decision maker’s preference on the acceptable risk for deﬁnition of the initial cost, and found Pareto solutions using a multiobjective genetic algorithm. Let T and X denote the service life of the structure and the vector of design variables, respectively. The annual occurrence rate of major seismic events is denoted by ν, which is modeled by a Poisson process. The annual monetary discount rate is denoted by λ; i.e., the value of money is reduced every year at the ratio of λ. The cost for the jth seismic damage state and its occurrence probability are denoted by Cj and Pj , respectively. Then the total cost C(T, X) considering nd damage states is formulated as (Liu, Wen, and Burns 2004) n ∑ ν C(T, X) = (1 − e−λT ) Cj Pj λ j=1 d

(5.13)

198

Optimization of Finite Dimensional Structures

The cost function Cj is given as Cj = Cjdamage + Cjcontent + Cjrelocation + Cjeconomic + Cjinjury + Cjfatality (5.14) where Cjdamage is the direct structural/nonstructural damage and repair cost, Cjcontent is the cost due to loss of contents, Cjrelocation is the relocation cost, Cjeconomic is the direct/indirect economic loss, Cjinjury is the human injury cost, and Cjfatality is the human fatality cost. Kocer and Arora (1999) deﬁned the total initial cost CI of a transmission tower, as follows, as the sum of the costs of material, painting, galvanizing, and welding: CI = Csg CW W + CP Asurface + CG W + Cweld Lweld

(5.15)

where Csg is the coeﬃcient indicating the steel grade, CW is the cost per unit weight of the steel, W is the weight of steel, CP is the cost of painting per unit area, Asurface is the surface area to be painted, CG is the unit cost of galvanizing, Cweld is the cost of welding per unit length, and Lweld is the length to be welded. The welding cost Cweld includes the costs for labor, electrode, power, and equipment. In another formulation (Cheng 2002), the total cost CT is ﬁrst classiﬁed into the initial cost CI and the damage cost CD as CT = CI + CD

(5.16)

and CD is further classiﬁed into CD = Cr + Cc + Ce + Cs + Cf

(5.17)

where Cr is the repair or replacement cost of the structure, Cc is the loss of contents, Ce is the economic impact of structural damage, Cs is the cost of injuries caused by structural damage, Cf is the cost of fatalities from structural damage or collapse.

5.2 5.2.1

Local and global searches of approximate optimal designs Introduction

Approximate optimal solutions under stress constraints have been conventionally found using the stress-ratio approach of fully-stressed design (FSD); see Sec. 1.6 for the FSD of trusses. Recently, it has been pointed out that there may exist many FSDs of building frames with almost the same total structural volume (Mueller, Liu, and Burns 2002; Liu and Burns 2003). Therefore,

Optimization of Building Frames

199

obtaining only one solution will not be suﬃcient for practical purposes, where several solutions satisfying stress constraints should be compared in view of other performance measures, including requirements in construction and fabrication processes. Furthermore, the objective function need not be strictly minimized; i.e., it will be helpful for the designer if several approximate optimal solutions with diﬀerent distributions of cross-sectional areas are obtained. The nonuniqueness of the optimal solution is generally conceived as a negative aspect in structural optimization, because it deteriorates the convergence property of the optimization algorithm; however, the nonuniqueness of the solutions, which is classiﬁed as follows, can be extensively utilized to generate many diﬀerent approximate optimal solutions: (a) Local nonuniqueness due to independence of the displacements and the constraint functions on the design variables, which is enhanced by regularity of the frame. (b) Global nonuniqueness due to nonconvexity of the objective and/or constraint functions. Sim˜ oes (1989) studied isolated global optimality of the truss optimization problem using second-order optimality conditions. Similar nonuniqueness can be observed for a plate (sheet) discretized to ﬁnite elements and subjected to in-plane loads (Kutylowski 2002). Jog and Haber (1996) derived the conditions for the stability (uniqueness) of the optimal solution using an incremental form of the variational problem. They suggested that the nonuniqueness of the solution to a compliance optimization problem can be detected by the singular values of the matrix deﬁned by the derivatives of the optimality conditions and the equivalent nodal load vector with respect to the design variables. Barthold and Gerzen (2009) utilized singular value decomposition of the sensitivity matrix to reduce the number of variables and to utilize second-order sensitivity coeﬃcients in the optimization process. Petersson (1999) investigated convergence properties of the optimization process with respect to mesh size and noted that the objective function may be insensitive to the thickness variation for special simple loading conditions, such as uniaxial tension/compression and simple shear, if the element with a bilinear displacement interpolation function is used. Watada and Ohsaki (2009a) investigated the nonuniqueness of the solution as the branching of the solution path based on the so-called continuation method with respect to the problem parameter. The continuation method for the optimization problem is basically the same as the parametric programming approach (Gal 1979; Fiacco 1983) or the hom*otopy method (Shin, Haftka, Watson, and Plaut 1988; Watson and Haftka 1989) for tracing the optimal solutions corresponding to the various parameter values (Nakamura and Ohsaki 1988). In this section, three formulations are presented for local search of optimal solutions of a regular plane frame under an stress constraints. The distance

200

Optimization of Finite Dimensional Structures

between the solutions is then deﬁned, and the approximate optimal solutions are globally searched consecutively so as to maximize the distance from the already found approximate solutions under an upper-bound constraint on the total structural volume (Ohsaki 2006a, 2006b).

5.2.2

Optimization problem and optimality conditions

Consider a rigidly jointed regular plane frame, as shown in Fig. 5.1, in Sec. 5.1.2. Let A∗i denote the cross-sectional area of the ith member. The second moment of inertia Ii∗ and section modulus Zi∗ are deﬁned as continuous functions of A∗i (see Appendix A.8). Hence, A∗i can be considered as the only independent design variable of member i. The members are classiﬁed into groups based on the symmetry and regularity properties of the frame, and the number of variables is reduced by using the so-called design variable linking approach. Let Ai denote the crosssectional area of the members in the ith group, and the vector consisting of Ai of m groups is denoted by A = (A1 , . . . , Am )⊤ . Let P denote the static load vector that is supposed to be independent of A. The number of degrees of freedom is denoted by n. The nodal displacement vector U against P is found by solving the stiﬀness equation K(A)U = P

(5.18)

where K(A) is the n × n stiﬀness matrix, which is a function of A. Alternatively, for a given U, the equivalent nodal load vector F(U, A) is deﬁned by F(U, A) = K(A)U

(5.19)

For a frame modeled using the beam element, constraints are given for the stresses σai , σbi , σci , and σdi at the member ends ‘a’, ‘b’, ‘c’, and ‘d’, as shown in Fig. 5.2, of the members i = 1, . . . , m∗ , where m∗ is the number of members. The same upper bound σ U is given, for simplicity, for the absolute values of the stresses of all members. The conventional assumption of a rigid ﬂoor is used. Hence, there is no axial force in a beam, and the absolute values of the stresses satisfy |σai | = |σbi | and |σci | = |σdi |. Therefore, the number of points at which the stresses are constrained is two for beams and four for columns that undergo axial deformation in addition to bending deformation. Constraints are also given for the nodal displacements. Let Hj (U(A), A) denote the jth inequality constraint function representing the bound on stress or displacement. The total length of the members in the ith group is denoted by Li . The optimization problem for minimizing the

Optimization of Building Frames

201

total structural volume V (A) is formulated as Minimize

V (A) =

m ∑

Ai Li

(5.20a)

i=1

subject to Hj (U(A), A) ≤ 0, (j = 1, . . . , nI ) ALi ≤ Ai ≤ AU i , (i = 1, . . . , m)

(5.20b) (5.20c)

where ALi and AU i are the lower and upper bounds for Ai , respectively, and nI is the number of inequality constraints. Problem (5.20) is a nonlinear programming (NLP) problem that can be solved using a gradient-based optimization algorithm; see Appendix A.2.2. Since Problem (5.20) is generally nonconvex, convergence to the global optimal solution is not guaranteed. However, if solutions with the same objective value are always found starting with diﬀerent initial solutions, it is highly possible the solutions are globally optimal. Even if the solution is not globally optimal, it satisﬁes the Karush-Kuhn-Tucker (KKT) conditions that are the necessary conditions for local optimality (see Appendix A.2.2.3 for details). Let µ = (µ1 , . . . , µnI )⊤ (≥ 0)denote the vector of Lagrange multipliers for the inequality constraints. The constraint function with respect to A only is e j (A) = Hj (U(A), A); i.e., the displacements U are considered denoted as H as implicit functions of A. If we consider the side constraints separately from the general inequality constraints, the Lagrangian S(A, µ) of Problem (5.20) is deﬁned as m nI ∑ ∑ e j (A) S(A, µ) = Ai Li + µj H (5.21) i=1

j=1

The derivative Si of S with respect to Ai is given as n ∑ I

Si = Li +

µj

j=1

where

ej ∂H , (i = 1, . . . , m) ∂Ai

∑ ∂Hj ∂Uk ej ∂H ∂Hj = + ∂Ai ∂Uk ∂Ai ∂Ai

(5.22)

n

(5.23)

k=1

The KKT conditions Si = 0 Si ≥ 0 Si ≤ 0

are written as for ALi < Ai < AU i , for Ai = ALi , for Ai =

(i = 1, . . . , m)

(5.24)

AU i

with inequality conditions and complementarity conditions µj ≥ 0, Hj ≤ 0, µj Hj = 0, (j = 1, . . . , nI )

(5.25)

202

Optimization of Finite Dimensional Structures

If A is locally optimal, there exist U and µ that satisfy the KKT conditions (5.24) and (5.25). Let J E denote the set of indices of member groups satisfying ALi < Ai < U Ai . Eq. (5.24) for i ∈ J E can be written as n ∑ I

Li +

j=1

µj

ej ∂H = 0, (i ∈ J E ) ∂Ai

(5.26)

The number of groups in J E is denoted by nE ; i.e., we have nE equations (5.26). Let C E and sE denote the set of indices of independent active constraints in (5.20b) and the number of constraints in C E , respectively; i.e., we have sE unknown nonnegative Lagrange multipliers µj , and the remaining multipliers are 0. If the jth and kth constraints are active and dependent, i.e., ej ek ∂H ∂H =c , (i ∈ J E ) (5.27) ∂Ai ∂Ai with c being a constant, the kth constraint is eliminated and (1 + c)µj is conceived as an independent multiplier in (5.26). Note that the Lagrange multipliers are available from the output of the NLP algorithms, such as sequential quadratic programming. However, even if µj is not available from the optimization results, µj can be computed from (5.26) after obtaining A and U of the optimal solution and carrying out design e j /∂Ai , because the existence of µj ≥ 0 sensitivity analysis to compute ∂ H E (j ∈ C ) is assured from the local optimality of the solution. In the following, we consider only the nondegenerate case in which µj is uniquely determined and the following conditions are satisﬁed:

5.2.3 5.2.3.1

Si > 0 for Ai = ALi

(5.28a)

Si < 0 for Ai = AU i µj > 0 for Hj = 0

(5.28b) (5.28c)

Local search of approximate optimal solutions Approximate condition for a sandwich section

Consider the frame consisting of members with sandwich sections or solid sections with constant heights; i.e., the second moment of inertia and section modulus are proportional to the cross-sectional area. Let AE denote the vector consisting of the cross-sectional areas of the members in groups in J E , and denote by ∆AE the increment of AE . If there is a direction ∆AE that does not have any eﬀect on U, then Hj , including displacement constraints and stress constraints, do not explicitly depend on AE , and the optimal solution is locally nonunique. Since P is independent of AE , the incremental form of

Optimization of Building Frames

203

(5.18) for variation ∆AE of AE and the corresponding variation ∆U of U is written as K(AE )∆U + KA (U)∆AE = 0 (5.29) A where the (i, j)-component Kij (U) of the n × nE matrix KA (U, AE ), which is called the stiﬀness matrix with respect to cross-sectional areas, is given using (5.19) as ∂Fi (U, AE ) A Kij (U, AE ) = (5.30) ∂AE j

The nonuniqueness of the optimal solution can be detected, as follows, using (5.29) on the basis of the rank of KA . Since Ii∗ is proportional to A∗i for each member with a sandwich section, KA does not explicitly depend on AE . Let an n-vector fi denote the nodal load vector representing the pair of loads equivalent to the nodal forces of member i and applied to the nodes connected to member i. Then the jth column of KA is equal to the sum of (1/A∗i )fi of the members in the jth group. Let ωi (ω1 ≥ ω2 ≥ · · · ) denote the ith singular value of KA , and deﬁne the diagonal rectangular matrix Ω so that its (i, i)-component is equal to ωi and the remaining components are 0. Then singular value decomposition (SVD) of KA leads to KA = QΩR⊤ (5.31) where the ith columns of the n × n matrix Q and the nE × nE matrix R are the left and right singular vectors Qi and Ri , respectively, corresponding to ωi , and the following relation is satisﬁed (see Appendix A.1.3 for details of SVD): KA Ri = ωi Qi , (i = 1, . . . , nE ) (5.32) It is assumed here that nE is less than n, which is usually satisﬁed by a rigidly jointed regular frame even if the design variables are not linked and the cross-sectional areas of all the members are considered to be independent design variables. If KA is full column rank, i.e., rank KA = nE , it can be observed from (5.29) that there is no solution ∆AE (̸= 0) satisfying ∆U = 0 except for the case where there exists a member that is not deformed; however, such member has a lower-bound cross-sectional area and is not included in the set J E at the optimal solution. For a nonoptimal and nonregular general frame, the rank of KA is generally equal to nE , because variation of AE in any pattern will lead to variation of U. However, it often happens for a regular frame that rank KA is less than nE . Suppose rank KA = nE − 1; i.e., ωnE = 0. In this case, (5.32) leads to KA RnE = 0

(5.33)

Therefore, ∆U = 0 is satisﬁed from (5.29) for ∆AE = RnE ; hence, U is constant if AE is modiﬁed in the direction of RnE . The constraints are satisﬁed by the modiﬁed design, and there exist µj > 0 (j ∈ J ) satisfying the

204

Optimization of Finite Dimensional Structures

KKT conditions, if Hj does not contain explicit term with respect to AE and e j are continuous with respect to AE . These the sensitivity coeﬃcients of H conditions are rigorously satisﬁed for the stress constraints of a frame with a sandwich section. Therefore, in this case, the optimal solution is nonunique with ﬁrst-order approximation and ∆AE satisﬁes the condition n ∑ E

E ∆AE i Li = 0

(5.34)

i=1 E where LE i is the sum of the length of members in the ith group in J . This way, optimal solutions can be locally searched using SVD of KA for a frame consisting of members with sandwich sections. Even if ωnE in (5.32) is not equal to 0 and has a very small positive value compared to the maximum singular value ω1 , approximate optimal solutions can be found by searching AE in the direction of RnE . For optimal frames with a general cross-section under constraints that explicitly depend on the cross-sectional areas, the independence of U on AE may be regarded as a ﬁrst-order approximate condition for local nonuniqueness.

5.2.3.2

Approximate condition for a general cross-section

More rigorous conditions for nonuniqueness of optimal frames with general cross-sections can be derived by relaxing the requirement of constant displacements. Suppose the optimal solution is fully stressed; i.e., sE ≥ nE . Let G denote the nI × m sensitivity matrix for which the (j, i)-component Gji is equal to the sensitivity coeﬃcient of the jth constraint function with respect to the ith design variable: ej ∂H Gij = (5.35) ∂Ai Suppose the constraints are numbered such that the ﬁrst sE constraints are active; i.e., the jth component of the vector HE of active constraints is deﬁned as HjE = Hj (i = 1, . . . , sE ). The reduced sE × nE sensitivity matrix corresponding to the independent set of active constraints and the variables in the set J E only is denoted by GE . Then, the ﬁrst-order approximation ∆HE of the incremental vector of active constraints HE is given as ∆HE = GE ∆AE

(5.36)

Therefore, the search direction of the approximate optimal solutions satisfying the constraints and optimality conditions (5.26) can be found by the SVD of the reduced sensitivity matrix GE . Note that the solution is locally nonunique if the singular value ωnE is 0. In this case, HjE are insensitive to the variation of AE in the direction of the nE th right singular vector of GE . The E E approximate Lagrange multipliers µE j corresponding to Hj (i = 1, . . . , s ) for

Optimization of Building Frames

205

the solution after increment can be found from (5.26) assuming the continuity of the optimal solution and its sensitivity coeﬃcients. This way, the SVD of sensitivity matrix can be eﬀectively used for local search of the optimal solution of a frame with a general cross-section. 5.2.3.3

General conditions for nonuniqueness

General conditions for local nonuniqueness without restriction on the crosssectional types and the numbers of sE and nE can be obtained by diﬀerentiating all the state equations and KKT conditions in a similar manner as the parametric programming approach in Appendix A.5. Let t denote an auxiliary parameter that deﬁnes the problem without any physical meaning. The A (i, j)-components of K and KA are denoted by Kij and Kij , respectively. E E The derivatives of µj , Ai , and U of the optimal solution with respect to the parameter t are found from the following equations, which are derived by diﬀerentiating (5.18), (5.20), and (5.26): n ∑

dUj ∑ A dAE dPi j Kij + = , (i = 1, . . . , n) Kij dt dt dt j=1 j=1

(5.37a)

n n ∑ ∂Hj dUi ∑ ∂Hj dAE k + = 0, (j = 1, . . . , sE ) E dt ∂U dt ∂A i k i=1

(5.37b)

nE

E

k=1

s ∑ s n e E dAE ∑ ∑ e j dµE ∂2H ∂H j j k + = 0, (i = 1, . . . , nE ) (5.37c) µE j E E E dt dt ∂A ∂A ∂A i i k j=1 j=1 k=1 E

E

E

e E are derived using the relation (5.23). where the sensitivity coeﬃcients of H j Let X denote the vector consisting of the unknown variables U, AE i , and . Then, from (5.37), the derivative of X with respect to t is obtained by µE j solving the linear equations in the following form: Z

dX =z dt

(5.38)

where Z and z are the constant matrix and vector, respectively. If Z is singular, then there exists a vector dX/dt corresponding to z = 0. Therefore, the direction dAE i /dt of the nonunique solutions can be found from the singular vector corresponding to the zero singular value of Z. Note that similar conditions can be obtained from the second-order optimality conditions (Sim˜oes 1989). The condition above is more rigorous than the previous approximate conditions for nonuniqueness, because it utilizes the exact diﬀerentiation of the governing equations and KKT conditions. However, in this approach, the Hessian of HjE is needed, which is very diﬃcult to obtain in the usual process of optimization.

206

Optimization of Finite Dimensional Structures 1

2 (1)

FIGURE 5.6:

5.2.4

3 (2)

4 (3)

5 (4)

6

7

(5)

(6)

A six-span continuous beam.

Global search of approximate optimal solutions

Because Problem (5.20) is a nonconvex problem, there may exist many local optimal solutions that have slightly larger objective values than that of the global optimal solution. In view of practical application, the most preferred solution should be chosen from a set of approximate solutions in view of other performance measures, including constructability and stiﬀness, against other sets of design loads. Let Ve denote the optimal objective value of Problem (5.20), and assign the following requirement as an approximate optimal solution: V (A) ≤ Ve + ∆V

(5.39)

where ∆V is assumed to be suﬃciently small. Then the approximate solub (k) tions are successively found by solving the optimization problems. Let A denote the kth approximate optimal solution that has already been found. b (k) is deﬁned by the Euclidean norm as The distance D(k) between A and A v um u∑ (k) b(k) )2 (5.40) D = t (Ai − A i i=1

An auxiliary variable τ is introduced, and the following optimization problem is solved for maximizing the minimum distance from existing optimal solutions under constraints on the responses and the total structural volume: Maximize τ

v um u∑ b(k) )2 , (k = 1, . . . , q) subject to τ ≤ t (Ai − A i

(5.41a)

i=1

V (A) ≤ Vb + ∆V

(5.41b)

Hj (U(A), A) ≤ 0, (j = 1, . . . , n ) I

ALi

≤ Ai ≤

AU i ,

(i = 1, . . . , m)

(5.41c) (5.41d)

where q is the number of global and approximate solutions that have already been found. This way, approximate optimal solutions with various distributions of cross-sectional areas can be found consecutively by solving Problem (5.41).

Optimization of Building Frames

207

Example 5.2 Nonuniqueness of the optimal solution is ﬁrst investigated for a continuous beam consisting of m members (elements) and m + 1 nodes subjected to concentrated moment M0 at each node. An example for m = 6 is shown in Fig. 5.6, where the numbers with and without parentheses are member numbers and node numbers, respectively. In the following examples, the units of force and length are kN and mm, respectively, which are omitted for brevity. The length L of each element is 2000, the elastic modulus E is 200, and the upper bound σ for the absolute value of stress is 0.07. The lower bound ALi for the cross-sectional area is 100, whereas the upper bound is not given. The concentrated nodal moment is 10,000. Design variable linking is not used; i.e., the number of variables is m. The beam has a sandwich section satisfying Ii = h2 Ai

(5.42)

where h = 50 is the distance between each ﬂange and the center axis, which is the same for all members. Constraints are given for the stresses of the ﬂanges at the two ends of each member. Let θi denote the rotation of the ith node. To investigate the special case of nonunique optimal solution, we assign a periodic boundary condition such that θ1 = θm+1 . As a small example, consider the case m = 2. Using the periodic boundary condition, and assuming Ai > ALi (i = 1, 2), i.e., J E = {1, 2}, the matrices K and KA are written as ( ) ( ) 4 2 4θ1 + 2θ2 4θ1 + 2θ2 A K = c(A1 + A2 ) , K =c (5.43) 2 4 2θ1 + 4θ2 2θ1 + 4θ2 where ci = Eh/L. Therefore, it is seen from the SVD of KA that U = (θ1 , θ2 )⊤ is insensitive to the design modiﬁcation ∆A = (1, −1)⊤ in the direction of the right singular vector corresponding to zero singular value, and the absolute values of the stresses of the ﬂanges of members 1 and 2 are equal to σ, which satisﬁes M0 σ= (5.44) h(A1 + A2 ) Therefore, the sensitivity matrix is obtained as ( ) M0 1 1 G=− h(A1 + A2 )2 1 1

(5.45)

Hence, the singular vector corresponding to the zero singular value is obtained as ∆A = (1, −1)⊤ after normalization so that the maximum absolute value is 1, which is the same as the vector obtained by the SVD of KA . Consider next the case m = 6 to investigate nonuniqueness through computational optimization. Optimization is carried out by IDESIGN Ver. 3.5 (Arora and Tseng 1987), where sequential quadratic programming is used. A

208

Optimization of Finite Dimensional Structures

TABLE 5.4:

Optimization results of the continuous beam

with m = 6. A1 A2 V (×106 )

ξ=0 714.286 714.286 8.57143

ξ = 0.5 717.170 711.401 8.57143

ξ = 1.0 719.964 708.607 8.57143

ξ = 2.0 724.229 704.343 8.57143

uniform random number 0 ≤ Ri < 1 is generated to deﬁne the initial solution as Ai = A0 (1 + ξRi ), (i = 1, . . . , 6) (5.46) where ξ is a parameter, and A0 = 100. Table 5.4 shows the optimization results from various initial solutions generated by diﬀerent values of ξ, where the solutions are periodic such that A1 = A3 = A5 and A2 = A4 = A6 . Note that each optimal solution has been found with eight iterations of sequential quadratic programming and satisﬁes the stress constraint with equality at one of the ﬂanges at an end of each member; i.e., the optimal solutions are fully stressed. Therefore, we can assume that the global optimal solution has been successfully found for all cases, although the optimization problem is nonconvex. The objective values are the same for all cases, although the cross-sectional areas are diﬀerent. Hence, the optimal solution is nonunique. Note that the nodal rotations have the same value 4.6667 × 10−3 at all nodes of all solutions; i.e., the displacement vector is independent of the cross-sectional areas of the optimal solutions, and the optimal solutions can be locally searched using the SVD of KA . The rank of KA is 5, and the singular values are 14.000, 12.124, 12.124, 7.0000, 7.0000, 0.0. The singular vector R6 corresponding to the vanishing singular value is (−1, 1, −1, 1, −1, 1)⊤ after normalization. Therefore, the optimal solutions can be written with a parameter α as A1 = A3 = A5 = 714.286 + α, A2 = A4 = A6 = 714.286 − α

(5.47)

and have the same objective value, which agrees with the result in Table 5.4. If the requirement θ1 = θ7 of the periodic boundary condition is not given, the rank of KA of the optimal solution is 6, and the singular values are 12.064, 11.151, 9.1058, 8.3791, 5.3656, 2.7534. Therefore, the rank deﬁciency of KA is strongly related to the regularity of the frame.

5.2.5 5.2.5.1

Numerical example of a regular plane frame Description of the frame model

Optimal solutions are locally and globally searched for a six-story six-span frame, as shown in Fig. 5.1, where H = W = 4000, m∗ = 84, and n = 97. The

Optimization of Building Frames

TABLE 5.5:

209

Optimization results of the six-story six-span

frame. A1 A2 A3 A4 V (×109 )

ξ=0 22,047.3 14,026.0 8266.1 4474.1 2.4824

ξ = 0.5 22,041.2 13,992.8 8265.3 4486.2 2.4824

ξ = 1.0 22,040.5 14,008.7 8274.1 4473.1 2.4824

ξ = 2.0 22,036.4 13,983.8 8271.7 4484.5 2.4824

members are classiﬁed into 45 groups, i.e., m = 45, considering the symmetry condition. Note that the units of force and length are kN and mm also in this section. Only horizontal loads are applied, and (P1 , P2 , P3 , P4 , P5 , P6 ) = (50, 100, 150, 200, 250, 300) in Fig. 5.1. The elastic modulus E is 200 and the upper-bound stress σ is 0.07. For the example of a sandwich section, h = 250 in (5.42) and ALi = 3000 for all groups. For the frame with wide-ﬂange beams and box columns, ALi = 5000 for all groups, and the following relations are assumed for the second moment of inertia Ii and section modulus Zi of members in the ith group so that only the cross-sectional areas are the design variables (Sawada and Matsuo 2003): { Columns : Ii = 1.076(Ai )2 , Zi = 0.804(Ai )1.5 (5.48) Beams : Ii = 3.648(Ai )2 , Zi = 1.580(Ai )1.5 The initial solutions are randomly generated in the same manner as (5.46) with A0 = ALi . 5.2.5.2

Local search

First we optimize the frame with a sandwich section under stress constraints only. The optimization results of Problem (5.20) are shown in Table 5.5, where the optimal objective value and the cross-sectional areas of the external columns of stories 1–4 denoted by A1 , . . . , A4 are listed. As is seen, the optimal objective values are 2.4824×109 for all cases, although the cross-sectional areas are diﬀerent. Because the optimization from four diﬀerent initial solutions converged to the same objective value, we can assume that the global optimal solutions have been successfully obtained, although uniqueness of the solution is not satisﬁed. The optimal solution for ξ = 0, denoted by Solution 1, is shown in Fig. 5.7, where the width of each member is proportional to its cross-sectional area. The number nE of groups satisfying Ai > ALi is 36. The optimal solution is fully stressed even for the group with Ai = ALi ; i.e., the maximum absolute values of the stresses are equal to the upper bound σ in all groups. The number of active constraints at the optimal solution is 126, which is larger than nE for all cases. First, we carry out SVD for KA of Solution 1 in Fig. 5.7. The 20 lowest and the maximum singular values are listed in the second column of Table 5.6.

210

FIGURE 5.7:

Optimization of Finite Dimensional Structures

Solution 1 of the six-story six-span frame.

Since there is a jump between ω25 and ω24 , the number of nearly zero singular values is 12. Let σij denote the stress at the jth point (j = 1, . . . , 4) in the ith member where the stress is constrained as shown in Fig. 5.2. The accuracy of an approximate solution is conﬁrmed by the maximum stress ratio β deﬁned by { } |σij | β = max (5.49) i,j σ where β = 1 corresponds to a fully stressed design. The cross-sectional areas are parametrically varied in the direction of the right singular vector R36 corresponding to the smallest singular value. The unit increment ∆AE is deﬁned so that it is proportional to R36 and the maximum variation of crosssectional area is equal to 1. Then the cross-sectional areas are varied from the b E as A b E + η∆AE with cross-sectional parameter η. The optimal solution A b ratio of V to V of Solution 1 is denoted by γ. Variations of β and γ with respect to the cross-sectional parameter η are plotted in Fig. 5.8. For example, if η = 200, i.e., if we allow maximum variation 200 of Ai , then β = 1.0069 and γ = 1.0006; hence, the ratios of increase of the objective function and the maximum stress ratio are suﬃciently small compared with the variation of the cross-sectional areas. Note that the stress constraints are violated in some members, because the nodal displacements are not strictly constant for design variation in the direction of R36 . Optimal solutions are next found for the frame with wide-ﬂange beams and box columns. In addition to the stress constraints, the upper bound 120 is given for the horizontal displacement at the roof level, which corresponds to an average drift angle of 1/120. The number nE of groups satisfying Ai > ALi at the optimal solution is 36. The optimal solutions are fully stressed, and the displacement constraint is satisﬁed with equality; hence, the number of active constraints at the optimal solution is 127, which is larger than nE . The

Optimization of Building Frames

211

Singular values of KA for a sandwich section and G for wide-ﬂange beams and box columns.

TABLE 5.6:

E

ω36 ω35 ω34 ω33 ω32 ω31 ω30 ω29 ω28 ω27 ω26 ω25 ω24 ω23 ω22 ω21 ω20 ω19 ω18 ω17 ··· ω1

KA 1.0960 × 10−2 1.8110 × 10−2 1.9496 × 10−2 2.4163 × 10−2 2.9663 × 10−2 3.8929 × 10−2 4.3074 × 10−2 5.1079 × 10−2 6.1228 × 10−2 7.8563 × 10−2 1.0079 × 10−1 7.0597 × 10−1 19.538 27.132 31.412 33.566 37.562 44.657 47.874 54.960 ··· 124.01

GE 4.1060 × 10−6 8.6195 × 10−6 1.1614 × 10−5 1.3993 × 10−5 1.8806 × 10−5 2.3470 × 10−5 2.5060 × 10−5 3.3610 × 10−5 3.9091 × 10−5 4.1709 × 10−5 4.8968 × 10−5 5.1727 × 10−5 5.5647 × 10−5 6.2508 × 10−5 7.6444 × 10−5 7.8804 × 10−5 8.8555 × 10−5 9.8923 × 10−5 1.0774 × 10−4 1.1527 × 10−4 ··· 4.8069 × 10−4

results of the SVD of GA are listed in the third column of Table 5.6. As is seen, no clear zero singular value is observed in this case. Variation of β with respect to the cross-sectional parameter η is as shown in Fig. 5.9 for the cross-sectional variation in the direction of the right singular vector R36 corresponding to the smallest singular value ω36 . Note that γ is not shown because it is very close to 1, e.g., 1.00008 at η = 200. Hence, approximate solutions with good accuracy can be obtained utilizing the SVD of GE even for frames with general cross-sectional shapes. 5.2.5.3

Global search

Approximate optimal solutions are globally searched under an upper-bound constraint of V that is 2% larger than the optimal objective value Vb of the global optimal solution in Fig. 5.7; i.e., ∆V = 0.02Vb in (5.41b). Problem (5.41) has been solved to maximize the minimum distance from the existing approximate optimal solutions. The nine solutions indicated by Solutions 2–10, which are found consecutively, are shown in Figs. 5.10(a)–(i),

212

Optimization of Finite Dimensional Structures Stress ratio β and volume ratio γ

1.01 1.008 1.006

β 1.004 1.002

γ 1 0

20 40 60 80 100 120 140 160 180 200 Deviation of cross-sectional area η

FIGURE 5.8: Variation of the maximum stress ratio β (solid line) and the volume ratio γ (dashed line) with respect to cross-sectional parameter η using singular vector of KA for a sandwich section. 1.01

Stress ratio β

1.008 1.006 1.004

β

1.002 1 0

20 40 60 80 100 120 140 160 180 200 Deviation of cross-sectional area η

FIGURE 5.9: Variation of the maximum stress ratio β with respect to cross-sectional parameter η using singular vector of GE for wide-ﬂange beams and box columns. respectively. As is seen, approximate solutions with various distributions of cross-sectional areas have been successfully found. Note that the constraints on the total structural volume are satisﬁed with equality, i.e., V = 1.02Vb for all solutions. The values of the minimum distance τ for Solutions 2–10 are plotted in Fig. 5.11, which conﬁrms that τ decreases as more solutions are found. This way, approximate solutions can be globally searched. However, it is possible to incorporate directly the preference of the designer under constraint

Optimization of Building Frames

213

(a) Solution 2

(b) Solution 3

(d) Solution 5

(e) Solution 6

(f) Solution 7

(g) Solution 8

(h) Solution 9

(i) Solution 10

FIGURE 5.10:

(c) Solution 4

Globally searched approximate optimal solutions.

(5.41b) on the structural volume with ∆V = 0.02Vb . For example, if the variance of the cross-sectional areas is minimized, the cross-sectional areas shown in Fig. 5.12(a) are obtained. Fig. 5.12(b) shows the solution that minimizes the maximum cross-sectional area, which turns out to be 13878.8, which is about 60% of 22,047.3 of Solution 1.

214

Optimization of Finite Dimensional Structures

Minimum distance

4000

3000

2000

1000 2

4

6

8

10

Solution number

FIGURE 5.11:

Variation of minimum distance from the existing solutions.

(a)

(b)

FIGURE 5.12: Approximate optimal solutions obtained by direct consideration of the designer’s preference on cross-sectional areas; (a) minimum variance, (b) minimum maximum value.

5.3 5.3.1

Parametric optimization of frames Introduction

Most of the optimization methods developed for trusses and frames are veriﬁed for small-scale problems and may not always be applicable to large-scale problems due to diﬃculties in computational cost and convergence properties. To resolve these diﬃculties, some methods, e.g., multilevel decomposition, have been developed (Kirsch 1975; Friedman and Fuchs 1987). Two-level decomposition can be used in many ﬁelds of optimization, including resource allocation problems, min-max type multiobjective programming problems, and

Optimization of Building Frames

215

best approximation problems (Shimizu, Ishizuka, and Bard 1997). Decomposition of an optimization problem was ﬁrst studied for a linear programming problem (Dantzig and Wolfe 1960) and was applied to a plastic limit design problem (Woo and Schmit 1981; Kaneko and Ha 1983). Substructure approaches have also been developed (Kirsch 1972; Nguyen 1987; Svensson 1987); however, they utilize the speciﬁc properties of the structural models and the optimization problems; i.e., they are not developed in a general form. A general framework for solving optimization problems of complex systems is called multidisciplinary optimization (MDO), which was developed in the 1990s in the ﬁelds of mechanical engineering and aeronautical engineering for application to optimization problems considering responses in several different disciplines (Sobieszczanski-Sobieski and Haftka 1996). Note that the concept of MDO is diﬀerent from that of the conventional decomposition method for large-scale analysis, e.g., the substructuring method and the domain decomposition method (Smith, Bjørstad, and Gropp 1996). In MDO, the design process of a complex structural system is divided into subsystems based on the properties of responses, or disciplines, e.g., structural analysis, computational ﬂuid dynamics, and electromagnetic analysis. For example, in automobile design, the simple approach of decomposition is based on objects such as powertrain, body, chassis, and electronics, whereas it can be partitioned in view of disciplines, e.g., durability, packaging, dynamics, safety, and noise-vibration-harshness (Kim, Michelena, Papalambos, and Jiang 2003). An interaction between aerodynamic analysis and structural analysis is illustrated in Fig. 5.13. The pressure load as an input to structural analysis is obtained as an output of aerodynamic analysis, whereas the initial condition of the aerodynamic analysis is determined using the deformation obtained by structural analysis. The interactions among the diﬀerent subsystems are modeled using parameters that deﬁne the design problem of each subsystem. An optimization problem that has parameters in addition to variables is called a parametric programming problem (Gal 1979; Ant´onio 2002), which can be conceived as a general framework of an optimization method such as sequential unconstrained minimization techniques (SUMT) (Fiacco and Cormic 1968) and interior point methods (Kojima, Shindoh, and Hara 1997) (see Appendix A.5 for basic equations of parametric programming). Parametric programming approaches have been widely applied to structural optimization, where the sensitivity of the optimal design with respect to the problem parameter, called optimum design sensitivity (Barthelemy and Sobieszczanski-Sobieski 1983; Schmit and Chang 1984; SobieszczanskiSobieski, James, and Dovi 1985; Vanderplaats and Yoshida 1985), is computed to investigate the properties of the optimal solutions and also to compute the sensitivity of the solution of the lower-level problem of a multilevel optimization problem with respect to the solution of the upper-level problem that is conceived as a parameter for the lower-level problem (Vanderplaats 1988;

216

Optimization of Finite Dimensional Structures Aerodynamics

Structures Static load

Aerodynamic analysis Structural analysis

Pressure

Deformation

FIGURE 5.13: Illustration of a multidisciplinary system with aerodynamics and structures.

Bloebaum, Hajela, and Sobieszczanski-Sobieski 1992). Two-level approaches have been extensively studied in the ﬁeld of hybrid approaches to optimization with anti-optimization (Elishakoﬀ, Haftka, and Fang 1994; Elishakoﬀ and Ohsaki 2010), where optimal solutions are found considering the worst-case scenario of the uncertain parameters. A simple cycle-based approach can be used for this purpose (Gurav, Langhaar, Goosen, and van Keulen 2005). There have been many papers on multilevel optimization of frames (Ding and Esping 1991; Friedman and Fuchs 1987), where the system variables, e.g., the nodal locations, are optimized in the upper level, and the component variables, e.g., cross-sectional properties, are optimized in the lower level. Salajegheh (1996) used a dual method with approximation for two-level optimization of frames. Nakamura and Ohsaki (1988) presented a parametric programming approach for generating a set of optimal trusses under multiple frequency constraints. In this respect, the parametric programming approach can be generalized in the framework of the continuation method (Mittelmann and Roose 1990) and the hom*otopy method (Shin, Haftka, Watson, and Plaut 1988; Watson and Haftka 1989). Ohsaki and Arora (1993) presented a general parametric programming approach to structural optimization. In this section, we ﬁrst describe the simple two-level decomposition approach to the design of frames and then present a general framework of non-hierarchical decomposition, which is applied to multilevel optimization of building frames (Ohsaki 1997a; Ohsaki, Nagano, and Wakamatsu 2000).

5.3.2

Two-level decomposition of frames

The simplest MDO approach is the two-level decomposition of the problem. Let xi = (xi1 , . . . , ximx )⊤ denote the vector of global or system-level design variables of the ith member, representing, e.g., the cross-sectional area and second moment of inertia, where mx is the number of global variables related to each member that is assumed to be the same for all members. The vector of all the global variables consisting of xi (i = 1, . . . , m) is denoted by X =

Optimization of Building Frames

217

(X1 , . . . , Xmg )⊤ , where m is the number of members, and mg = m × mx is the total number of global variables. The vector of local or member-level design variables of the ith member is i ⊤ denoted by yi = (y1i , . . . , ym y ) , which consists of the height, width, etc., of y the cross-section, where m is the number of local variables for each member that is also assumed to be the same for all members. Suppose the global variables are deﬁned by the local variables as xi = f (yi ), (i = 1, . . . , m)

(5.50)

Note that the same vector f = (f1 , . . . , fmx )⊤ is used for all members, for simplicity. If nx < ny , then the optimal value of yi may be found through member-level optimization for a speciﬁed value of xi as a result of global optimization. If nx = ny , then yi is computed directly from xi = f (yi ) without carrying out optimization. If nx > ny , then there might exist no feasible value of yi satisfying xi = f (yi ). Let U denote the vector of nodal displacements that are conceived as the state variables. The vector of nodal displacements of the ith member is denoted by ui . The constraints on state variables, e.g., on nodal displacements in the global optimization problem, are generally given with inequalities Hj (X) ≤ 0 (j = 1, . . . , nH ), where nH is the number of global constraints. Note that the state variables are implicitly included as a function of X in the deﬁnition of Hj (X), and the derivatives of Hj (X) are computed using the design sensitivity analysis presented in Chap. 2. The local constraints, e.g., the stress constraints, are given as hij (xi , ui (X), yi ) ≤ 0 (j = 1, . . . , nh ), where nh is the number of local constraints that is the same for all members, and ui is a function of X. Only inequality constraints are considered, for simplicity. Consider the following optimization problem for minimizing an objective function F (X): Minimize

F (X)

(5.51a)

subject to Hj (X) ≤ 0, (j = 1, . . . , n ) H

hij (xi , ui (X), yi ) XiL ≤ Xi ≤ XiU , i i

(5.51b)

≤ 0, (j = 1, . . . , n ; i = 1, . . . , m) h

(i = 1, . . . , mg )

x = f (y ), (i = 1, . . . , m) yjiL

≤

yji

≤

yjiU ,

(5.51c) (5.51d) (5.51e)

y

(j = 1, . . . , n ; i = 1, . . . , m)

(5.51f)

where the superscripts ( · )U and ( · )L denote the upper and lower bounds, respectively. The optimal solution of Problem (5.51) is to be found by iteratively solving the upper-level (global or system level) and lower-level (local or member level) problems. There are many approaches to the deﬁnition of these problems (Sobieszczanski-Sobieski, James, and Dovi 1985). The most serious diﬃculty for solving the lower-level problem is that it often happens that the

218

Optimization of Finite Dimensional Structures

problem is infeasible due to unrealistic value of the parameters deﬁned by the variables in the upper-level problem. One possible approach to avoid infeasibility is to maximize margin, or minimize violation, of the constraints instead of satisfying all the constraints. Another approach is to deﬁne the objective function C i (hi ) of the lower-level problem of the ith member as the penalty for violating the local constraints; i.e., the objective function of the lowerlevel problem is not related to the objective function F (X) of the upper-level problem. Hence, cumulative penalty function C i (hi ), deﬁned as follows, may be minimized in each member: n ∑ h

i

i

C (h ) =

Cji (hij )

(5.52)

j=1

where Cji (hij ) is deﬁned so that Cji (hij ) > 0 for hij > 0, and Cji (hij ) = 0 for hij ≤ 0 (see Appendix A.2.2.4 for details of penalty function approaches). Note that C i (hi ) should be a continuously diﬀerentiable and non-decreasing function of hi . Then the lower-level problem for the ith member is formulated as Minimize

C i (hi ) i

(5.53a) i

subject to x = f (y )

(5.53b)

hij (ui (X), yi ) ≤ yjiL ≤ yji ≤ yjiU ,

h

0, (j = 1, . . . , n ) y

(j = 1, . . . , n )

(5.53c) (5.53d)

where yi is the variable vector, and xi and ui are conceived as parameter vectors that are speciﬁed as the result of upper-level optimization. Therefore, the optimal solution and the corresponding optimal objective value of the lower-level problem are found for each speciﬁed value of xi and ui . Hence, the optimal solution and the optimal objective value may be conceived as functions of xi and ui , consequently, as functions of X, which are denoted by e i (X), respectively. Then the sensitivity coeﬃcients of y e i (X) and C e i (X) and y i e C (X) with respect to X, which are called parametric sensitivity coeﬃcients, can be found using the techniques of parametric programming or optimum design sensitivity analysis, as described in Appendix A.5. The values of the solution at the kth iteration of the upper-level problem for optimizing X are e i (X) is linearly approximated as denoted by the superscript ( · )(k) . Then, C ( ei

ei

C (X) = C (X

(k)

)+

ei ∂C ∂X

)⊤ (X − X(k) )

(5.54)

e i (X). Hence, the upper-level optimization Similar approximation is given for y

Optimization of Building Frames

219

1

A1 , I1 2

A2 , I 2

P1 , U1

P2 , U 2

FIGURE 5.14:

A two-bar frame.

problem is formulated as Minimize

F (X)

(5.55a)

subject to Hj (X) ≤ 0, (j = 1, . . . , n ) e i (X) ≤ 0, (i = 1, . . . , m) C H

XiL yjiL

≤ Xi ≤ ≤

yeji (X)

XiU , ≤

(5.55b) (5.55c) g

(i = 1, . . . , m )

yjiU ,

(5.55d) y

(j = 1, . . . , n ; i = 1, . . . , m)

(5.55e)

In the optimization process, the initial value X(0) is ﬁrst given for X, and the state variables U are computed. Then xi and ui are transmitted to the lower-level problems as parameters, and the optimal local variables as well as the Lagrange multipliers are found. Then the upper-level problem is solved, where the sensitivities of optimal variables and objective value are eﬀectively incorporated to improve the convergence property. Example 5.3 An example is given for the deﬁnition of variables and constraints for a rigidly jointed frame that consists of two members, as shown in Fig. 5.14. The relation among the variables, parameters, and constraints of the global system (frame) and the subsystems (members) are illustrated in Fig. 5.15. Suppose the two members have rectangular cross-sections, and let ai and bi (i = 1, 2) denote the height and width, respectively, of the section of the ith member. The cross-sectional area and second moment of inertia of member i are denoted by Ai and Ii , respectively, which are functions of ai and bi . The global variables are x1 = (A1 , I1 )⊤ and x2 = (A2 , I2 )⊤ , which are combined as X = (A1 , I1 , A2 , I2 )⊤ . The global constraints are given for the nodal displacements U1 and U2 , which are functions of X; i.e., H1 (X) = U1 (X) − U1U , H2 (X) = U1L − U1 (X), H3 (X) = U2 (X) − U2U , H4 (X) = U2L − U2 (X)

(5.56)

220

Optimization of Finite Dimensional Structures Frame Global constraints

U 1L≤ (A1 , I 1 , A 2 , I 2 ) ≤ U1U U 2L≤ (A1 , I 1 , A 2 , I 2 ) ≤ U2U

Member 1

Member 2

Local constraints σ 1L ≤ σ1( a1 , b 1 ) ≤ σ1U

A1 = f 1 (a1 , b1) I 1 = f 2 (a1 , b1)

FIGURE 5.15: frame.

Local constraints σ 2L ≤ σ2( a2 , b 2 ) ≤ σ2U

A2 = f 1 (a2 , b2) I 2 = f 2 (a2 , b2)

Illustration of two-level decomposition of the two-bar

where upper and lower bounds are denoted by ( · )U and ( · )L , respectively. The global variables are deﬁned with respect to the local variables as A1 = f1 (a1 , b1 ), I1 = f2 (a1 , b1 ), A2 = f1 (a2 , b2 ), I2 = f2 (a2 , b2 )

(5.57)

Local constraints are given for the maximum absolute value σi of the stress (i = 1, 2), which is regarded as a function of ai and bi by ﬁxing the displacements ui as parameters. Let σiL and σiU denote the lower and upper bounds for σi , respectively. Then the local constraints are formulated as follows using the variable vectors y1 = (a1 , b1 )⊤ and y2 = (a2 , b2 )⊤ : h11 (u1 , y1 ) = σ1 (u1 , y1 ) − σ1U , h12 (u1 , y1 ) = σ1L − σ1 (u1 , y1 ), h21 (u2 , y2 ) = σ2 (u2 , y2 ) − σ2U , h22 (u2 , y2 ) = σ2L − σ2 (u2 , y2 )

5.3.3

(5.58)

General concept of decomposition to subsystems

The simplest approach to decomposition of large complex systems into subsystems is a hierarchical decomposition, as we have seen in the previous section. Kirsch (1975) deﬁned coordinate variables between the ﬁrst (lower) level and the second (upper) level problems, and demonstrated optimization of crosssectional areas and nodal coordinates of a small truss, where the nodal coordinates are used as coordination variables. The general concept of coordination of subsystems was presented by Mesarovi´c, Macko, and Takahara (1970). If the subsystems are optimized individually, it is very diﬃcult to satisfy feasibility in the system-level (global) constraints. Therefore, only the local constraints are to be satisﬁed when optimizing the subsystems. The feasibility of the local solution with ﬁxed global variables is called single discipline

Optimization of Building Frames

221

feasibility, and a solution satisfying feasibility in every discipline (subsystem) is said to have individual discipline feasibility (Cramer, Dennis, Frank, Lewis, and Shubin 1994). Kim, Michelena, Papalambos, and Jiang (2003) proposed target cascading for decomposing the optimization problem of a complex system. Consider a problem of minimizing the deviation of the responses from their target values. Let tij denote the target parameter vector transmitted from the jth subsystem in the ith level, while its response is denoted by rij . Then consistency conditions are formulated with respect to the diﬀerence vector cij = tij − rij , and the penalty for violation of consistency is added in the objective function for optimizing the subsystem with some appropriate equality and inequality constraints. Tosserams, Etman, Papalambos, and Rooda (2006) extended the target cascading approach to utilize the augmented Lagrangian method. It should be noted that most of the decomposition approaches are heuristic and convergence of the iterative procedure is not proved. Even if the solution converges, the convergence property is not very good if the dependencies among the subsystems are highly nonlinear and global optimality of the global system is not guaranteed. If the system is separable, then the variable vector is partitioned to the vectors yi of local variables without any duplication, and the objective function F (y1 , . . . , ym ) is formulated as the sum of the objective functions Fi (yi ) of subsystems as m ∑ F (y1 , . . . , ym ) = Fi (yi ) (5.59) i=1

and the constraints of the ith subsystem are deﬁned using yi only. In this case, the optimal solution of the global system is found successfully by optimizing each subsystem. If the objective function includes the term deﬁned by the global variables, then the constraints with respect to the global variables are to be satisﬁed. Haftka and Watson (2005) deﬁned quasiseparable subsystems and formulated the optimization problem as Minimize

F (X, y1 , . . . , ym ) = F0 (X) +

m ∑

Fi (yi )

(5.60a)

i=1

subject to Hj (X) ≤ 0, (j = 1, . . . , nH ) hij (yi )

≤ 0, (j = 1, . . . , n ; i = 1, . . . , m) h

(5.60b) (5.60c)

In this case, if the lower-level problem is formulated to maximize the margin of the local constraints, then the lower-level problem is always feasible, and the optimal solution can be obtained by iteratively solving the upper- and lower-level problems. However, for general complex systems, it is not always possible to decompose the system hierarchically to a tree-type structure of the subsystems, because the subsystems are highly likely to interact with each other. There-

222

Optimization of Finite Dimensional Structures

SG C pG

xG

Sk

Sj pj C

FIGURE 5.16: subsystems.

xk

j

x

j

pk

Ck

Relation between parameters of a global system and two

fore, a non-hierarchical system is required to model general systems (Balling and Sobieszczanski-Sobieski 1996).

5.3.4

Parametric multidisciplinary optimization problem

In this section, we generalize the concept of two-level parametric optimization to problems consisting of hierarchical and non-hierarchical subsystems. Suppose a large structural system is divided into several subsystems, and optimization of the global system is carried out iteratively optimizing the subsystems. In order to improve the convergence property, we can utilize parametric programming approaches, as discussed in the previous section. The global system S is divided into nS subsystems S1 , . . . , SnS . Generally, each pair of subsystems has interacting parameters and variables, and the parameters for Sj will be modiﬁed after optimizing another subsystem Sk . Consider ﬁrst a hierarchical system, where nS subsystems exist below the global system. A simple case of nS = 2 is illustrated in Fig. 5.16, which is a generalization of the two-bar frame in Example 5.3. Let xG and pG denote the vectors of design variables and parameters of the global system, for which the optimization problem is formally written as Minimize

C(xG , pG )

(5.61a)

subject to x ∈ X (p ) G

where X G is the feasible region for xG .

G

G

(5.61b)

Optimization of Building Frames

223

Let xj = (xj1 , . . . , xjmj )⊤ denote the vector of mj design variables of subsystem Sj . The objective function of Sj is given as Cj (xj , pj ), where pj = (pj1 , . . . , pjnP )⊤ is the vector of nP parameters. The problem to be solved in j

the jth subsystem is simply written as Minimize

Cj (xj , pj )

subject to x ∈ Xj (p ) j

j

(5.62a) (5.62b)

where Xj is the feasible region for xj . Note that Cj (xj , pj ) may be the same as C(xG , pG ), or may be the penalty function for the local constraints, as discussed in the previous section. Suppose pG and pj (j = 1, . . . , nS ) are functions of xj (j = 1, . . . , nS ) G ⊤ is conceived and xG , respectively. Then, the vector pG = (pG 1 , . . . , pnP ) G G G e (x ) through optimization of as an implicit function of x denoted by p the subsystems, and the objective function with respect to the vector xG = G ⊤ (xG 1 , . . . , xmG ) only is deﬁned as e G ) = C(xG , p e G (xG )) C(x

(5.63)

e with respect to xG is written as Accordingly, the sensitivity coeﬃcient of C i ∑ ∂C ∂ peG e ∂C ∂C k = + , (i = 1, . . . , mG ) G ∂xG ∂xG ∂xG ∂p i i i k k=1 nP

(5.64)

j S G If pG k is an explicit function of x (j = 1, . . . , n ), the sensitivity of pk with respect to xj is easily computed. Furthermore, sensitivity of the optimal value of xj with respect to pj is found using the parametric programming approach, and the dependence of pj on xG is also assumed to be known. G Hence, the sensitivity coeﬃcient ∂ peG k /∂xi can be easily obtained. By utilizing G e the sensitivity coeﬃcient ∂ C/∂xi in (5.64), instead of using only ∂C/∂xG i in the right-hand side of (5.64), convergence of a gradient-based optimization process may be improved. Next, we consider non-hierarchical systems. Suppose, for simple presentation of formulations, that we have two subsystems j and k. Fig. 5.17 illustrates the dependencies of the variables and parameters between two subsystems. The parameter vector pj of Sj is a function of xk , and pk is a function of xj . If we use the simple cyclic approach, pk (xj ) is ﬁrst computed by assigning xj , and Sk is to be optimized to ﬁnd the optimal value of xk . Because the optimal solution of Sk is found for each speciﬁed value of pk , it is a function of pk (xj ), ek (xj ). Then pj is updated and accordingly, is a function of xj denoted as x j k from p (x ), and the subsystem Sj is optimized. Hence, the parameter pj e j (xj ), and the may be conceived as an implicit function of xj denoted by p j j j j j j e (x ) = C (x , p e (xj )). objective function of x only is deﬁned as C

224

Optimization of Finite Dimensional Structures Sj Cj

Sk Ck

xj

xk

pj

pk

FIGURE 5.17: Relation between parameters of two non-hierarchical subsystems. e j (xj ) considering the variation of the paThe sensitivity coeﬃcients of C rameters of the subsystems with respect to xji is obtained for general nonhierarchical systems as ej ∂C ∂xji

=

∂C j ∂xji

nP

+

j ∑ ∂C k ∂ pejr

∂pjr ∂xji nP j mk k j j ∑ j ∑ xs ∂C ∂ per ∂e ∂C , (i = 1, . . . , mj ) = + j j k ∂x ∂xji ∂p ∂x s r s=1 i r=1 r=1

(5.65)

The relation (5.65) is utilized to improve the convergence property of optimization of subsystems, which may be carried out sequentially or simultaneously. The second-order sensitivity coeﬃcients can be incorporated, if necessary, as presented in Appendix A.5.

5.3.5 5.3.5.1

Optimization of plane frames Problem formulation

In this section, a parametric programming approach is presented for optimization of plane frames, which are divided into subsystems based on analysis types, i.e., static analysis and dynamic analysis (Ohsaki 1997a). The static analysis is carried out using the frame model consisting of beams and columns, while the dynamic analysis is carried out using the shear model. As Subsystem 1, we formulate the optimization problem of the shear model under constraint on the eigenvalue of vibration. The objective function C D (D) is the total structural volume, which is a function of the vector of story stiﬀnesses D = (D1 , . . . , DnF )⊤ , where nF is the number of stories. The lower bounds of the lowest eigenvalue Ω1 (D) and Di are denoted by ΩL and DiL , respectively. Then the optimization problem of the shear model is formulated

Optimization of Building Frames

225

as Minimize

C D (D, pD )

subject to Ω1 (D) ≥ Ω Di ≥

DiL ,

(5.66a) L

(5.66b) F

(i = 1, . . . , n )

(5.66c)

where pD is the parameter vector deﬁning the relation between the structural volume and story stiﬀnesses, which can be obtained from the stiﬀnesses of the members as the solution of the optimization problem of Subsystem 2 described below. Stress constraints under the speciﬁed static loads are considered in the optimization problem of Subsystem 2 with the cross-sectional areas A = (A1 , . . . , Am )⊤ as design variables. The length of the ith member is denoted by Li . The constraints are given for the stresses σij at the four points (j = 1, . . . , 4) of the ith member (i = 1, . . . , m), as described in Fig. 5.2, and their upper and lower bounds are denoted as σiU and σiL , respectively. Let ∆i (A) and Qi denote the interstory drift and shear force of the ith story under static horizontal loads. Then, the stiﬀness Di of the ith story is deﬁned as Di = Qi /∆i (i = 1, . . . , nF ), and the optimization problem for minimizing the total structural volume C A (A) is formulated as

Minimize

C A (A) =

m ∑

Ai Li

(5.67a)

i=1

subject to σiL ≤ σij (A) ≤ σiU , (i = 1, . . . , m; j = 1, . . . , 4) Qi F = pA i , (i = 1, . . . , n ) ∆i (A) Ai ≥ ALi

(5.67b) (5.67c) (5.67d)

where pA i = Di is the speciﬁed value of the stiﬀness of the ith story, which is the optimal solution of Subsystem 1, and is conceived as the parameter of Subsystem 2. It is easily seen from (A.122) in Appendix A.5 that the parametric sensitivity coeﬃcients of the objective function of Problem (5.67) with respect to pA i (= Di ) are equal to the Lagrange multipliers λi for the constraints (5.67c). The second-order sensitivity coeﬃcients of the objective function, denoted by e A /∂pA ∂pA , are found from (A.124). Let the superscript ( · )(k) Wij = ∂ 2 C i j denote the value at the kth iteration. Since C D is the total structural volume, which is an implicit function of D, the parameter vector pD consists of λj and Wij , and the objective function C D (D, pD ) of Problem (5.66) is approximated

226

Optimization of Finite Dimensional Structures

as follows using pA i = Di : n ∑ F

e D (D) ≃ C D (D(k) ) + C

(k)

λi (Di − Di )

i=1

+

1 2

nF ∑ nF ∑

(k)

(k)

Wij (Di − Di )(Dj − Dj )

(5.68)

i=1 j=1 n ∑ F

+τ

(k)

(Di − Di )4

i=1

where τ is the parameter for preventing divergence in the iterative process. A large value for τ leads to a large penalty for increments of Di . This penalty is not needed if the matrix W = (Wij ) is positive deﬁnite. The optimization algorithm is summarized as follows: Step 1: Assign the coeﬃcients τ , λi , and Wij for (5.68) and the initial value (0) Di for Di . Set the iteration counter k = 0. (k)

Step 2: Solve Problem (5.66) to obtain the optimal value of Di by assigning the parameters pD for the deﬁnition of the objective function C D (D, pD ). (k)

A(k)

Step 3: Incorporate the optimal value of Di to pi , and solve Problem (k) (5.67) to obtain the optimal value of Ai . Then compute the ﬁrst- and second-order parametric sensitivity coeﬃcients λi and Wij , respectively, of the objective function with respect to pD(k) = D(k) . Step 4: Update the iteration counter as k ← k + 1 and go to Step 2 if not converged. 5.3.5.2

Numerical examples

Optimal solutions are found for a six-story three-span plane frame with base beams, as shown in Fig. 5.18, subjected to horizontal loads representing 2 seismic loads. The elastic modulus is 205.8 kN/mm , the bounds for stress 2 are σiU = −σiL = 68.6 N/mm , and ALi = 5.0 × 10−3 m2 for all members. The mass of each story is 4.0 × 104 kg. The horizontal loads (kN) for Problem (5.67) are given as (P1 , . . . , P6 ) = (31.578, 45.469, 60.168, 76.442, 96.785, 159.96). Note that the self-weight is not considered. (0) The initial values are Di = 98.0 kN/m and λi = 1.0 m4 /N for all stories, 5 2 and Wij = 1.0 m /N for i = j and 0 for i ̸= j. It is conﬁrmed that the optimal solution does not depend on the initial solution. The parameter τ is ﬁxed at 1.0412 × 10−2 mm5 /N2 . Each member consists of a sandwich section, for simplicity, where the distance between the ﬂanges is 0.6 m. The number of

Optimization of Building Frames

227

P6 4m P5 4m P4 4m P3 4m P2 4m P1 4m

8m

FIGURE 5.18:

8m

8m

A six-story three-span frame.

Optimal cross-sectional areas Ai (×104 mm2 ). Floor/Story Ext. beam Int. beam Ext. column Int. column 1 2.6761 0.5000 2.8547 3.1243 2 4.9633 1.4121 2.3583 3.2674 3 3.7934 3.0018 1.5722 3.5504 4 2.5645 4.0827 1.0598 3.1192 5 1.6073 3.9255 0.5733 2.4236 6 1.0577 2.6093 0.5000 1.2400 7 0.5000 0.7432

TABLE 5.7:

design variables is 26 considering the symmetry conditions. The optimization software package IDESIGN (Arora and Tseng 1987) is used. The lower bound for the fundamental eigenvalue is 40.0 rad2 /s2 . The optimal cross-sectional areas are obtained as shown in Table 5.7. Note that the beams in the ﬁrst and seventh ﬂoor means the base beams and those in the roof level, respectively. The optimal story stiﬀnesses D1 , . . . , D6 (×108 N/m) are 3.4730, 3.2573, 2.9363, 2.4114, 1.7521, 1.0322, and the optimal objective value is 6.0994 m3 , which is close to 6.0956 m3 obtained without using the parametric programming approach. The number of cycles in Steps 2–4 is 2, and the CPU time is 59.6% of that without the parametric programming approach.

228

Optimization of Finite Dimensional Structures y 1

2

3

4

A B C

x

D

(a) Plan

(b) Plane frame (x-dir.)

(c) Plane frame (y-dir.)

(d) Lumped-mass shear model

FIGURE 5.19:

5.3.6 5.3.6.1

A 4 × 4 span eight-story three-dimensional frame.

Optimization of a three-dimensional frame Deﬁnition of seismic load and story stiﬀness

Consider a three-dimensional regular frame, as shown in Fig. 5.19, which is divided into plane frames, as shown in Figs. 5.19(b) and (c), in the xand y-directions, respectively. The optimal solution that minimizes the total structural volume is found under constraints on the responses against the seismic motions in the two horizontal directions. In the process of evaluating seismic responses, the plane frame is further simpliﬁed into a lumped-mass shear model, as shown in Fig. 5.19(d). The conventional assumption of a rigid ﬂoor is used. Let D∗ = (D1∗ , . . . , Dn∗ F )⊤ denote the vector of story stiﬀnesses in x- or ydirection of the three-dimensional frame, which are the design variables of the shear model. The story mass includes the non-structural mass and the structural mass, which is a function of the cross-sectional areas of members. The story stiﬀness of the plane frame is to be found so that the mean-maximum interstory drift δi (D∗ ) for the speciﬁed displacement response spectrum SD (Ωr ) evaluated by the square-root-of-sum-of-squares (SRSS) method is equal to the

Optimization of Building Frames

229

δi (D∗ ) = δ i , (i = 1, . . . , nF )

(5.69)

speciﬁed value δ i :

The details of the response spectrum approach are shown in Appendix A.7. Because the numbers of design variables Di∗ and the constraints (5.69) are the same for the shear model, the values of Di∗ are easily determined from (5.69) using any optimization technique, including the stress-ratio approach for fully-stressed design (see Sec. 1.6). After obtaining Di∗ , the equivalent static story shear force Q∗i and the horizontal load Pi∗ to be applied to the plane frame in order to optimize the member cross-sectional areas are deﬁned as follows in terms of the interstory drift and the story stiﬀness: Q∗i (D∗ ) = Di∗ (D∗ )δ i , (i = 1, . . . , nF ) Pi∗ (D∗ ) = Q∗i+1 (D∗ ) − Q∗i (D∗ ), (i = 1, . . . , nF − 1)

(5.70a) (5.70b)

Pn∗F (D∗ ) = Q∗nF (D∗ )

(5.70c)

In the following, the argument D∗ is not written explicitly for simple presentation of the formulations. 5.3.6.2

Optimization for speciﬁed story stiﬀness

The cross-sectional areas of members are optimized under constraints on the story stiﬀness and the stresses against the speciﬁed static loads. Let a∗ and aj denote the total area of each ﬂoor and the area covered by the jth plane frame in the x- or y-direction. In the following, the value for the jth j plane frame is indicated with the superscript j. Then the story stiﬀness Di to be speciﬁed, horizontal load Pij , and story shear force Qji of the ith story of the jth plane frame are given as j

Di =

aj ∗ aj ∗ aj ∗ j j D , P = P , Q = Q i i a∗ i a∗ i a∗ i

(5.71)

In this case, the plane frames in each direction deform without interaction between them. Let Aji , Iij (Aji ), and Zij (Aji ) denote the cross-sectional area, second moment of inertia, and section modulus of the ith member of the jth plane frame. Note that Iij (Aji ) and Zij (Aji ) are functions of Aji . Two load vectors PSj and PHj , corresponding to the vertical live load and self-weight, and the horizontal loads that represent the seismic loads, respectively, are considered. The stress σij of the ith member of the jth plane frame is deﬁned as the maximum absolute value of the stresses at the two edges of two ends against PSj and PHj that are applied simultaneously. The set of loads PSj and −PHj is also applied if the frame is not symmetric. The bending stresses due to the distributed vertical loads along the beams are also considered; however, axial force does not exist in the beams because of the assumption of a rigid ﬂoor.

230

Optimization of Finite Dimensional Structures

The interstory drift of the ith story of the jth frame consisting of mj members is denoted by δij (Aj ), which is a function of Aj = (Aj1 , . . . , Ajmj )⊤ . Then the story stiﬀness Dij (Aj ) of the jth frame is calculated from Dij (Aj ) =

Qji , (i = 1, . . . , nF ) δi (Aj )

(5.72)

Let σiUj denote the upper bound for the maximum stress σij (Aj ) of the ith plane frame, which is assumed to be positive. The lower bound for Aji and the j length of the ith member are denoted by ALj i and Li , respectively. Then the problem for minimizing the total structural volume C j (Aj ) of the jth plane frame is formulated as m ∑ j

Minimize subject to

5.3.6.3

C j (Aj ) =

Aji Lji

i=1 j j σi (A ) ≤ σiUj , (i = 1, . . . , mj ) j Dij (Aj ) = Di , (i = 1, . . . , nF ) j Aji ≥ ALj i , (i = 1, . . . , m )

(5.73a) (5.73b) (5.73c) (5.73d)

Successive optimization of plane frames

A three-dimensional frame is optimized by successively optimizing the plane frames in the two directions. A serious diﬃculty arises, however, from the fact that the columns belong to two plane frames in diﬀerent directions; hence, the cross-sectional properties of the columns cannot be modiﬁed independently in the process of optimizing a plane frame. Therefore, the cross-sectional areas of the columns are ﬁxed in the lowerlevel problem for optimizing the beams of the plane frames. After optimization of all the plane frames is completed, the cross-sectional areas of the columns are modiﬁed using the parametric sensitivity coeﬃcients of the optimal solutions of the plane frames with respect to the cross-sectional areas of the columns, which are regarded as the parameters for the optimization problem of a plane frame. Then the beams of the plane frames are optimized for the updated values of the cross-sectional areas of the columns. Suppose the member numbers are assigned so that the members 1, . . . , mjb are beams and mjb + 1, . . . , mj are columns. The vector of cross-sectional areas of beams is given as Ajb = (Aj1 , . . . , Ajmj )⊤ . Let λj = (λj1 , . . . , λjmj )⊤ , b

j ⊤ µj = (µj1 , . . . , µjnF )⊤ , and η j = (η1j , . . . , ηm denote the vectors of nonj) b

negative Lagrange multipliers for the constraints (5.73b), (5.73c), and (5.73d), respectively. Then, the Lagrangian for Problem (5.73a) with ﬁxed cross-

Optimization of Building Frames

231

sectional areas of columns is deﬁned as j

ψ

j

(Ajb , λ, µ, η)

=

mb ∑

m ∑ j

Aji Lji

i=1

+

λji (σij (Ajb ) − σiUj )

i=1

n ∑

+

(5.74)

j

F

j

µji (Dij (Ajb ) − Di ) +

mb ∑

j ηij (ALj i − Ai )

i=1

i=1

The multipliers of the optimal solution are usually available if a gradientbased nonlinear programming approach is used. Even if the multipliers are not available, they are easily calculated from the optimality conditions after the solution has converged; see Sec. 5.2. The member numbers are assigned also for the three-dimensional frame so that the members 1, . . . , mb are beams and mb + 1, . . . , m are columns, where m is the total number of members. In the following, the values for the three-dimensional frame are indicated without superscript j. From (A.122) in Appendix A.5, the sensitivity coeﬃcients of the objective function C of the three-dimensional frame with respect to the cross-sectional areas Ak (k = mb + 1, . . . , m) of the columns, which are the same as the crosssectional areas Ajk (k = mjb + 1, . . . , mj ) of the columns of the jth frame, are calculated from j

mb n ∑ n ∑ n ∑ ∑ ∂C ∂σij ∂Dij = Lk + + , (k = mb + 1, . . . , m) (5.75) λji µji ∂Ak ∂Ak j=1 i=1 ∂Ak j=1 i=1 S

S

F

where nS is the number of plane frames. The optimization algorithm is summarized as follows: Step 1: Optimize each plane frame considering the cross-sectional areas of beams and columns as independent design variables. Deﬁne Ai for each column, e.g., as the mean or maximum value of the optimal crosssectional area of the corresponding column in the plane frames in two diﬀerent directions. Skip this process if default initial values are given for the cross-sectional areas of columns. Step 2: Find optimal cross-sectional areas of beams by optimizing each plane frame for the ﬁxed cross-sectional areas of columns, and evaluate the Lagrangian multipliers for the constraints. Note that the optimization of plane frames can be carried out in parallel. Step 3: Calculate the parametric sensitivity coeﬃcients of the objective function with respect to Ai of the columns using (5.75), and modify Ai based on the steepest descent method. Reduce the move limit ∆A for preventing divergence to β∆A (0 < β < 1) at each iteration. If there is no

232

Optimization of Finite Dimensional Structures feasible solution, modify Ai for a column as follows for each constraint that is violated: Interstory drift: Ai = γ1 Ai (γ1 > 1) Stress of beam: Ai = γ2 Ai (γ2 < 1) Stress of column: Ai = γ3 Ai (γ3 > 1)

(5.76a) (5.76b) (5.76c)

where γ1 , γ2 , and γ3 are the speciﬁed parameters. Note that if the stress constraint of a beam is violated, it can be satisﬁed by reducing the cross-sectional area of the columns connected to the beam so that the bending deformation of the beam is reduced. Step 4: Go to Step 2 if the solution is not converged. 5.3.6.4

Numerical examples

An optimum design is found for a 27-story three-dimensional frame with a plan view, as shown in Fig. 5.19(a). The span lengths (m) in the xand y-directions are (W12 , W23 , W34 ) = (12, 10, 10) and (WAB , WBC , WCD ) = (8, 8, 10), respectively, where, e.g., W12 is the distance between lines 1 and 2. 2 The elastic modulus is 205.8 kN/mm , the mass at each ﬂoor is 500 kg/m2 , and the mass density of the steel beams and columns is 7.86×103 kg/m3 . The beams and columns consist of wide-ﬂange and box sections, and the values of Iij and Zij are deﬁned as functions of Aji as Column : Iij = 1.2(Aji )2 , Zij = 1.0(Aji )1.5 Beam :

Iij

=

8.0(Ajj )2 ,

Zij

=

2.0(Aji )1.5

(5.77a) (5.77b)

The lower bound for the cross-sectional area is 0.05 m2 . The response spectrum by Newmark and Hall (1982) is used for representing seismic motions; see its deﬁnition and parameter values in Appendix A.7. The six lowest modes are considered for evaluating the seismic responses of the shear model. Stiﬀness-proportional damping is used with the damping ratio 0.02 for the lowest mode. The columns of the three-dimensional frame and the beams of each plane frame are classiﬁed, respectively, into 15 and 5 groups with the same crosssectional areas; see Ohsaki, Nagano, and Wakamatsu (2000) for details. The method of modiﬁed feasible directions in the library DOT 5.0 (VR&D 1999) is used for optimization. The upper bounds for the maximum interstory drift 2 and stresses are 0.02 m and 323.4 N/mm , respectively. The parameters for modiﬁcation of the cross-sectional areas of the columns are γ1 = 1.15, γ2 = 1.05, γ3 = 1.25, and β = 0.5. The cross-sectional area of each column at the ﬁrst stage is deﬁned to be equal to the larger value in the optimal plane frames in two directions. The optimal solution at the sixth step for ∆A = 0.02 m2 is as shown in Fig. 5.20,

Optimization of Building Frames

(A)

(1)

FIGURE 5.20: frame.

(B)

(2)

(C)

(3)

233

(D)

(4)

Optimization result of the 27-story three-dimensional

where (A)–(D) and (1)–(4) in the ﬁgure are deﬁned in Fig. 5.18, and the width of each member is proportional to the cross-sectional area. It may be observed from Fig. 5.20 that the columns in the center span have larger cross-sectional areas than the exterior columns.

234

5.4 5.4.1

Optimization of Finite Dimensional Structures

Local search for multiobjective optimization of frames Introduction

As we have discussed in the previous sections, the cross-sectional properties of the frames are usually selected from a list or catalog of the available standard sections. Therefore, the optimization problems are formulated as a combinatorial problem with continuous objective function, constraint functions, and state variables (see Sec. 1.3.2 for general formulation and classiﬁcation of structural optimization problems with discrete variables). It is very easy to solve a combinatorial optimization problem by enumerating all the possible combinations of the variables if the number of variables is small. However, the computational cost increases as an exponential function of problem size represented by numbers of variables and constraints; therefore, it is not possible to solve a practical problem using the enumeration approach within a practically admissible computational time. An important aspect of the practical design process is that it may be sufﬁcient to obtain an approximate optimal design rather than obtaining the global optimal solution. Another aspect of practical design optimization is that multiple performance measures are to be considered, and the problem turns out to be a multiobjective programming (MOP) problem. As deﬁned in Appendix A.4.1, the approaches to MOP are classiﬁed into those with and without a priori information of preference. In the latter approach, which is also called a method with posteriori articulation of preferences (Marler and Arora 2004), a set of Pareto optimal solutions is ﬁrst generated, and the most preferred solution is selected from the set by the decision maker. If we regard the solutions that are dominated by some other feasible solutions are not acceptable as candidate designs, the number of acceptable solutions is simply reduced by generating the set of Pareto optimal solutions. Therefore, in this approach, it is important that the Pareto optimal solutions with enough number and diversity be generated, or preferably all the Pareto solutions are enumerated. Heuristics have been extensively applied to obtain Pareto solutions of combinatorial MOP problems in many ﬁelds of engineering, including structural design (Jones, Mirrazavi, and Tamiz 2002). However, an eﬃcient optimization approach is desired for a structural optimization problem, because computationally expensive structural analysis should be carried out for evaluation of the objective and/or constraint functions. Therefore, in this section, we concentrate on simulated annealing and tabu search, which utilize local search and are categorized as single-point-search heuristics. Application of populationbased approaches, including the genetic algorithm (GA) and particle swarm optimization (PSO) to MOP problems are discussed in various books and papers (Goldberg 1989; Coello Coello, Pulido, and Lechuga 2004; Coello Coello

Optimization of Building Frames

235

and Pulido 2005). Ray and Liew (2002) utilized Pareto dominance to maintain diversity of the solutions, and used multilevel sieve to handle constraints for PSO. As an extension of GA, an immune algorithm (IA) (f*ckuda, Mori, and Tsukiyama 1999) can also be applied to MOP. Luh and Chueh (2004) presented an IA for constrained MOP of trusses. In this section, we ﬁrst present several approaches of single-point-search heuristics to MOP problems. Then, a simple local search method is presented for combinatorial multiobjective structural optimization (Ohsaki 2008) (see Appendix A.3 for details of single-point-search heuristic approaches to singleobjective problems and Appendix A.4 for various methodologies of MOP problems).

5.4.2 5.4.2.1

Heuristic approaches to combinatorial multiobjective programming Problem formulation and basic algorithm

Suppose the following list Xi is given as a set of ri available values for the variable xi : Xi = {xi1 , . . . , xiri }

(5.78)

Let J = (J1 , . . . , Jm )⊤ denote the vector of integer variables, where m is the number of variables. The jth value xij in Xi is assigned to xi if Ji = j. Consider, for simplicity, an unconstrained problem, and let Fi (J) denote the ith objective function to be minimized. Then an MOP problem with nF objective functions is formulated as minimize

F1 (J), . . . , FnF (J)

subject to Ji ∈ {1, . . . , ri }, (i = 1, . . . , m)

(5.79a) (5.79b)

Problem (5.79) is classiﬁed as a combinatorial MOP problem because it has integer variables and multiple objective functions. In a single-point-search heuristic approach, only one solution, called the seed solution, is obtained at each step of iteration. Let J(k) denote the seed solution at the kth step. A solution J∗ in the neighborhood of J(k) is accepted or rejected as the seed solution of the next step in accordance with each speciﬁc algorithm; see Appendix A.3 for details. 5.4.2.2

Simulated annealing

Simulated annealing (SA) is very eﬀective in problems with many local optimal solutions, because it allows a move to a non-improving solution. Although SA was originally developed for single-objective problems, it has been recently extended to MOP problems.

236

Optimization of Finite Dimensional Structures

Let J(k) denote the solution at the kth step of iteration. For a neighborhood solution J∗ , deﬁne ∆F as F n α ∑ Fi (J∗ ) − Fi (J(k) ) ∆F = F (5.80) n δi i=1 where α is the scaling parameter, and δi is a possible range of Fi for conversion of the ith objective value to a nondimensional value. A single Pareto optimal solution can be found if ∆F is used for the criterion for acceptance of the neighborhood solution J∗ in the same manner as the SA for a single-objective problem. Ho, Yang, Wong, and Ni (2003) used the ﬁtness function, which is similar to the Pareto ranking approach of a multiobjective GA (Goldberg 1989). A sharing function approach that is also used for multiobjective GA (Goldberg and Richerdson 1987) is incorporated to generate the Pareto solutions with enough diversity. The main drawback of sharing function for MOP is that a dominated solution tends to have large ﬁtness because many Pareto solutions are stored in the list and a new Pareto solution will have a large penalty with the sharing function. Ho, Yang, Wong, and Ni (2003) proposed an eﬃcient ﬁtness function to avoid this drawback. Viana and de Sousa (2000) applied a multiobjective SA to project scheduling using a weighted Chebychev metric. Jilla and Miller (2001) compared several methods of multiobjective SA and presented the following algorithm: Algorithm 5.1 Step 1: Randomly generate the initial solution J(0) , and set the temperature parameter as T = T0 . Initialize the Pareto candidate set as empty, assign δi for deﬁning the range of Fi , and set the iteration counter k = 0. Step 2: Randomly generate the neighborhood solution J∗ of J(k) . Step 3: If there is no solution in the Pareto candidate set that dominates J∗ , then accept J∗ as J(k+1) = J∗ and store it in the Pareto candidate set; otherwise, accept J∗ with the probability P deﬁned as ( ) ∆F P = exp − (5.81) Tk Add J(k+1) to the Pareto candidate list if J∗ has been accepted; otherwise, let J(k+1) = J(k) . Step 4: Decrease Tk to Tk+1 by the speciﬁed rule called the cooling schedule. In the most popular approach of the cooling schedule, Tk is updated to Tk+1 = ηTk with a speciﬁed constant η. Several steps of neighborhood search can be done at the same temperature.

Optimization of Building Frames

237

Step 5: If the termination condition is not satisﬁed, set k ← k + 1 and go to Step 2; otherwise, remove the dominated solutions in the Pareto candidate list, output the remaining Pareto optimal solutions, and terminate the process. This way, a set of Pareto solutions can be generated by conducting the algorithm only once. Czyzak and Jaszkiewicz (1998) presented a method called Pareto simulated annealing that maintains multiple solutions at each step like GA. The weights λj for the objective functions are updated through iteration to keep diversity of the solutions. The probability of acceptance of a neighborhood solution may be deﬁned using the Chebychev norm: { [ ( )]} λj (Fj (J(k) ) − Fj (J∗ )) P = min exp max (5.82) T (k) j=1,...,nF or the sum of the increments of the objective functions F ) n ( (k) ∗ ∑ λj (Fj (J ) − Fj (J )) P = min exp T (k) j=1

(5.83)

The solution is updated to J(k+1) that is closest to J(k) and is not dominated by J(k) . Then the weights are modiﬁed as κλ(k) if Fj (J(k+1) ) ≥ Fj (J(k) ) j (k+1) (5.84) λj = λ(k) /κ if F (J(k+1) ) < F (J(k) ) j j j where α (> 1) is a speciﬁed parameter. Whidborne, Gu, and Postlethwaite (1997) presented two approaches of SA to multiobjective control design. Their objective is not optimization but satisfaction of inequality constraints given as gj (J) ≤ εj (j = 1, . . . , nI ), where εj is the speciﬁed upper bound and nI is the number of constraints. Their ﬁrst method transforms the MOP problem to a problem with the single objective function F (J) using a goal programming based on a Chebychev norm as ( ) gj − εj F (J) = max ,0 (5.85) λj j=1,...,nI where λj is the weight parameter for the jth constraint. Their second algorithm utilizes Pareto optimality for deﬁning the acceptance probability of the solution. A function cj (J) is ﬁrst deﬁned as { εj if gj (J) ≤ εj cj (J) = (5.86) gj (J) if gj (J) > εj For two solutions J1 and J2 , the following three properties are deﬁned:

238

Optimization of Finite Dimensional Structures

P1: J1 is said to be superior to J2 if and only if cj (J1 ) ≤ cj (J2 ) for all j ∈ {1, . . . , n}. P2: J1 is said to be inferior to J2 if and only if cj (J1 ) ≥ cj (J2 ) for all j ∈ {1, . . . , n} and cj (J1 ) > cj (J2 ) for at least one j. P3: J1 is noninferior to J2 if and only if J1 is not either inferior or superior to J2 . If a neighborhood solution J∗ is superior to J(k) , then accept J∗ unconditionally as J(k+1) = J∗ . Otherwise, deﬁne a function φj as φj =

1 (cj (J∗ ) − cj (J(k) )) λj

(5.87)

and compute ∆F as the sum of φj for j such that cj (J∗ ) > cj (J(k) ). The probability of acceptance is given by ∆F using the standard process of SA. 5.4.2.3

Tabu search

Tabu search (TS), which is originally developed for single objective problems (Glover 1989), has also been shown to be eﬀective for MOP problems (Hansen 1997). In TS, the best solution in the neighborhood is selected as the next candidate, and a tabu list is used to prevent a local cyclic search among a small number of diﬀerent solutions. Armentano and Arroyo (2004) presented an algorithm for MOP that has multiple solutions at each step. Baykasoglu (2006) presented a multiobjective TS and applied it to simple structural optimization problems. The single-point-search algorithm for MOP by Baykasoglu, Owen, and Gindy (1999b) is summarized as follows: Algorithm 5.2 Step 1: Randomly generate the initial solution J(0) , which is chosen as the seed solution. Initialize the Pareto list P, tabu list T , and Pareto candidate list C as P = T = C = {J(0) }. Set the iteration counter as k = 0. Step 2: Generate the set of q neighborhood solutions N = {JN j | j = 1, . . . , q} (k) of the current seed solution J . Let S denote the set of candidate solutions in N that are not dominated by any solution in N , P, and C, and are not included in T . Step 3: Randomly select a solution J∗ from S. If S is empty, select the oldest solution in C as J∗ . Step 4: Remove the solutions in P and C that are dominated by a solution in S.

Optimization of Building Frames

239

Step 5: Add J∗ to P and T , and add other candidate solutions to C. Set J(k+1) = J∗ . Step 6: If C is empty and there exists no new candidate solution, or if the number of iterative steps exceeds the speciﬁed upper bound, terminate the process; otherwise, let k ← k + 1 and go to Step 2. The algorithm by Hansen (1997) maintains multiple solutions in a similar manner as the GA, and each solution has its own tabu list. Let δi denote the range of the ith objective function Fi , and deﬁne ci = 1/δi . The range parameter πi is deﬁned as ∑ ci πi = , c = ci c j=1 nF

(5.88)

The distance sj (J∗ ) from the solution J∗ to the set S is deﬁned using the Chebychev distance as sj (J∗ ) = min {max[λk (Fk (J∗ ) − Fk (Ji ))]} Ji ∈S

k

(5.89)

where λk is a weight coeﬃcient. The algorithm is summarized as follows: Algorithm 5.3 Step 1: Generate an initial set of solutions S, empty the tabu list, set the iteration counter k = 0, and initialize the weight coeﬃcients as λi = 1 (i = 1, . . . , nF ). Step 2: For each J in S, determine an appropriate search direction by updating λj in view of the relation between J and the Pareto optimal solutions in S, so that the Pareto solutions will be equally spaced in the objective function space. Then select the best seed solution J∗ , which is in the neighborhood of J and is not included in the tabu list. Step 3: Add J∗ to the tabu list and the Pareto set, and remove the solutions in the Pareto set that are dominated by J∗ . Replace a randomly selected solution in S with a randomly generated solution through the procedure called drift to maintain the diversity of Pareto solutions. Step 4: Go to Step 2 if the termination condition is not satisﬁed; otherwise, output a set of Pareto optimal solutions and terminate the process. Baykasoglu, Owen, and Gindy (1999a) presented a TS based on goal programming. It maintains only one solution at each step, and reaches a Pareto optimal solution; i.e., it cannot generate a Pareto optimal set with a single run. The algorithm is summarized as follows:

Optimization of Finite Dimensional Structures W1

P4

W2 10

1 1

5

2

8 13

5 14

3

8 15

6

12 m

FIGURE 5.21:

7 13

12

P1

7

4

2

W1

11

12

P2

W2

11 4

10 P3

W1

4@4m

P5

5m

240

9

6m

6m

12 m

A ﬁve-story four-span frame.

Algorithm 5.4 Step 1: Generate an initial solution J(0) , and regard J(0) as the best solution. Arrange the objective functions in nonincreasing order of priority; i.e., F1 (J) has the largest priority. Set the iteration counter k = 0. Step 2: Generate a set of neighborhood solutions of J(k) . Step 3: Evaluate the deviations of the objective functions from their goals in the objective function space, and determine the rank of the solutions based on the priorities of the objective functions. Step 4: Move to the best neighborhood solution J∗ that is not included in the tabu list. Step 5: Update the solution as J(k+1) = J∗ , and replace the best solution by J∗ if it is better than the previous best solution. Step 6: Construct the tabu list containing the recent best solutions. Step 7: Set k ← k + 1 and go to Step 2 if the termination condition is not satisﬁed.

5.4.3

Local search for multiobjective structural optimization

Consider a plane frame, as shown in Fig. 5.21, where the members are classiﬁed into m groups. Let J = (J1 , . . . , Jm )⊤ denote the vector of integer variables that represent the cross-sectional parameters. The jth value in the speciﬁed list of available sections is assigned to the ith group if Ji = j (see Sec. 5.4.2.1 and Appendix 5.4.2 for details of the problem formulation). If we use SA or TS for a multiobjective structural optimization problem, a set of Pareto optimal solutions can be found by carrying out structural analysis

Optimization of Building Frames

241

many times, say, several thousand times. However, for practical applications, the number of analyses should be reduced to several hundreds at the most. Furthermore, because we have many design variables, the performance of the algorithms applied to mathematical toy problems does not suggest any performance measure for a structural optimization problem. In this regard, the use of a tabu list in TS is not very eﬀective, and we may be able to generate some approximate Pareto solutions using a simple local search (LS), as described below. In order to show the eﬀectiveness of the local search in a simple manner, we consider a problem without constraints. The algorithm is an extension of Algorithm 5.2 in Sec. 5.4.2.3 for multiobjective TS by Baykasoglu, Owen, and Gindy (1999b). The objective functions Fi are normalized by the range parameter δi as Fi Fi∗ = , (i = 1, . . . , nF ) (5.90) δi Let F∗ and F∗(i) denote the objective values of the current seed solution and the ith solution in the set S of the candidate solutions, respectively. To obtain the Pareto optimal set with good accuracy and diversity, the following four strategies are used for the selection of the seed solution in Step 3 of Algorithm 5.2: Strategy 1: Randomly select a seed solution from S. Note that this strategy corresponds to Step 3 of the original algorithm. Strategy 2: Select the solution that minimizes

∑nF

∗(i) j=1 (Fj

∗(i)

Strategy 3: Select the solution that minimizes Fj jective function Fj .

− Fj∗ ).

− Fj∗ for a speciﬁc ob-

Strategy 4: Deﬁne the density φi of the solutions near Ji ∈ S in the set of Pareto solutions as ∑ φi = s(d(Ji , Jj )) (5.91) Jj ∈P

where d(Ji , Jj ) is the distance between Ji and Jj in the space of objective functions or design variables. In the following examples, the Euclidean distance in the objective function space is used. The sharing function s(d), which is often used in a multiobjective GA, is deﬁned as } { d s(d) = max 0, 1 − γ

(5.92)

where γ is the parameter called sharing radius or niche size. Then select the solution in S that has the smallest value of φi .

242

Optimization of Finite Dimensional Structures

5.4.4

Properties of Pareto optimal solutions

Let P denote the correct Pareto optimal set obtained by enumeration of all solutions. The size of approximate Pareto optimal set A obtained using a heuristic approach is denoted by nP . Various measures have been presented for evaluation of the properties of the approximate Pareto optimal set (Coello Coello, Pulido, and Lechuga 2004). We use the following measures: • Error ratio: Deﬁne ei so that ei = 0 if the solution i exists in the set P and ei = 1 if not. Then the error ratio ER is deﬁned as n 1 ∑ ER = A ei n i=1 A

(5.93)

The approximate set is included in the correct set if ER = 0. • Generational distance: Let di denote the Euclidean distance from the ith approximate solution to its nearest solution in P: v u nF u∑ ∗(i) ∗(j) di = min t (Fk − Fek )2 (5.94) j

k=1

∗(i) ∗(j) where Fk and Fek are the values of Fk∗ of the ith solution in A and the jth solution in P, respectively. Then, deﬁne the generational distance GD as v u nA ∑ 1 u (5.95) d2i GD = A t n i=1

The approximate set is included in the correct set if GD = 0. • Spacing: The minimum Manhattan distance gi between the ith approximate solution and a solution in P is deﬁned as n ∑ F

gi = min j

∗(i)

|Fk

∗(j) − Fek |

(5.96)

k=1

Let g denote the mean value of gi , and deﬁne the spacing SP as v u nA u 1 ∑ t SP = (g − gi )2 (5.97) n − 1 i=1 If the approximate solutions are uniformly distributed in the objective function space, then SP = 0.

Optimization of Building Frames

243

• Span: Span SN is deﬁned as the maximum distance between two approximate solutions: n ∑ A

SN =

∗(i)

max{|Fk i,j

k=1

∗(j)

− Fk

|}

(5.98)

If SN has a large value, then the solutions are widely distributed.

5.4.5 5.4.5.1

Numerical examples Mathematical problem

Consider ﬁrst the following problem with two objective functions with respect to four variables (Coello Coello, Pulido, and Lechuga 2004; Coello Coello and Pulido 2005): √ √ Minimize F1 (x) = 200(2x1 + 2x2 + x3 + x4 ) (5.99a) ( ) √ √ 1 2+2 2 2 2 2 F2 (x) = − + (5.99b) 100 x2 x3 x4 subject to 1 ≤ x1 ≤ 3 √ 2 ≤ x2 ≤ 3 √ 2 ≤ x3 ≤ 3 1 ≤ x4 ≤ 3

(5.99c) (5.99d) (5.99e) (5.99f)

The real variable xi is converted to an integer variable Ji by the following relation: xi (Ji ) = xLi + ∆xi × Ji

(5.100a)

− ti − 1

(5.100b)

∆xi =

xU i

xLi

L where xU i and xi are the upper and lower bounds of xi , and the number ti of the integer values is 214 = 16, 384 for x1 and x4 , and 213 = 8192 for x2 and x3 . In each of the following ﬁgures, the exact Pareto optimal set, called the Pareto front, obtained by enumeration is plotted in a solid line. The neighborhood solutions are generated using the random variable of normal distribution with 0 mean. The standard deviation is 800 for J1 and J4 , and 400 for J2 and J3 . Note that all the variables are modiﬁed simultaneously in the process of generating neighborhood solutions. If Ji < 1 or Ji > ti is satisﬁed as the result of updating Ji using the random number, Ji is replaced by 1 or ti , respectively. The range parameters of the objective values in (5.90) are δ1 = 600 and δ2 = 0.03. Fig. 5.22 shows the objective values of the 100 neighborhood solutions generated from a seed solution, indicated by the ﬁlled square. As is seen, the

244

Optimization of Finite Dimensional Structures 0.04 0.035 0.03

F2

0.025 0.02

Seed solution

0.015 0.01 0.005 0 1000

1200

1400

1600

1800

F1

FIGURE 5.22: Distribution of the neighborhood solutions of the mathematical problem.

TABLE 5.8:

Comparison of performances of the mathematical problem by TS and LS with various numbers of analyses. ER LS (8000) Minimum Maximum Mean Standard deviation TS (8000) Minimum Maximum Mean Standard deviation LS (400) Minimum Maximum Mean Standard deviation

GD (×10−3 ) SP (×10−2 )

SN

0.139 0.574 0.399 0.102

0.011 0.360 0.070 0.072

0.111 0.506 0.275 0.104

1.270 2.071 1.703 0.214

0.155 0.574 0.407 0.107

0.009 0.754 0.100 0.178

0.126 0.643 0.259 0.106

1.295 2.267 1.693 0.260

0.296 1.000 0.737 0.194

0.528 43.61 5.255 9.453

0.536 3.644 1.764 0.665

0.145 1.427 0.944 0.287

neighborhood solutions are uniformly distributed around the seed solution. Therefore, the Pareto front can be easily reached using Strategy 1 with random selection of the seed solution. The optimization results with various numbers of analyses are listed in Table 5.8, where the minimum, maximum, and mean as well as the standard deviation are computed with 30 trials from diﬀerent initial random seeds. The number of analyses for a single trial of each case is indicated in parentheses, e.g., LS (8000) means a local search with 8000 analyses. The number q of the

0.04

0.04

0.035

0.035

0.03

0.03

0.025

0.025

F2

F2

Optimization of Building Frames

0.02 0.015

245

0.02 0.015

0.01

0.01

0.005

0.005

0 1000

1200

1400

1600

0 1000

1800

1200

F1

1800

(b) Case 400-1-SN

0.04

0.04

0.035

0.035

0.03

0.03

0.025

0.025

F2

F2

1600

F1

(a) Case 8000-1-SN

0.02 0.015

0.02 0.015

0.01

0.01

0.005

0.005

0 1000

1200

1400

1600

0 1000

1800

1200

F1

1400

1600

1800

F1

(c) Case 400-1-GD

(d) Case 400-2-SN

0.04

0.04

0.035

0.035

0.03

0.03

0.025

0.025

F2

F2

1400

0.02 0.015

0.02 0.015

0.01

0.01

0.005

0.005

0 1000

1200

1400

1600

F1

(e) Case 400-3-F1-SN FIGURE 5.23:

1800

0 1000

1200

1400

1600

1800

F1

(f) Case 400-4-SN

Pareto solutions of the mathematical example.

neighborhood solutions is 10, which leads to 800 and 40 cycles for the cases of 8000 and 400 analyses, respectively, of updating a seed solution. Strategy 1 is used for selection of the seed solutions. The results of TS with the tabu list of length 1000 are also listed for comparison purposes. Since the average number of solutions that were rejected with the use of the tabu list with 8000 analyses is only 28, the use of the tabu list does not have any strong eﬀect on the results. Furthermore, about 40% of

246

Optimization of Finite Dimensional Structures

the solutions by TS and LS are included in the correct Pareto set. Therefore, in the following examples, we use LS without the tabu list. The results are presented by using an abbreviation based on the number of analyses, the strategy number for selection of the seed solution, and the performance measure used for selection of the results from the 30 trials; e.g., Case 400-1-SN indicates that the number of analyses is 400, and the Pareto set with maximum SN is selected from the 30 results by Strategy 1. For Strategy 3, the objective function to be minimized is indicated as Case 400-3-F1-SN, which shows that the solution with minimum F1 is to be selected. The ‘+’ marks in Fig. 5.23(a) are the approximate Pareto solutions of Case 8000-1-SN. Note that the Pareto solutions with good accuracy and diversity can be obtained by 8000 analyses in all 30 trials irrespective of the initial random seed. The performance measures for various cases are listed in Table 5.8. The result of 400 analyses with Strategy 1 with maximum SN is shown in Fig. 5.23(b). As is seen from Table 5.8, the four measures of 400 analyses are worse than those of 8000 analyses; however, it is observed from Fig. 5.23(b) that many diverse approximate Pareto solutions with good accuracy have been found within 400 function evaluations. The Pareto solutions for minimum GD are shown in Fig. 5.23(c), which shows that diversity of solutions is sacriﬁced if accuracy is increased. The result of maximum SN by Strategy 2 is shown in Fig. 5.23(d), which also has a small range of solutions. It may be observed from Figs. 5.23(b)–(d) that the solutions with small F1 may not be found if the two objective functions are equally evaluated. Therefore, we next minimize F1 using Strategy 3. If we maximize SN, the Pareto set obtained is shown in Fig. 5.23(e), which shows that the number of solutions in the small range of F2 has been sacriﬁced as a result of minimizing F1 . Finally, if we use Strategy 4 with the sharing function, the results of maximum SN are as shown in Fig. 5.23(f), where the radius δ is 0.1 for the normalized objective functions F1∗ and F2∗ . In this example, the use of the sharing function is very eﬀective in view of both accuracy and diversity. 5.4.5.2

Optimization of a plane frame

Consider a ﬁve-story four-span plane frame, as shown in Fig. 5.21, subjected to static loads, where the horizontal loads (kN) are (P1 , P2 , P3 , P4 , P5 ) = (7.5, 8.4, 10.1, 13.1, 31.5), and the vertical loads are W1 = 245 kN and W2 = 2 343 kN. The elastic modulus is 205.8 kN/mm . In the following example, the total structural volume and the compliance deﬁned in Sec. 1.8 are chosen as the two objective functions F1 (J) and F2 (J), respectively. The members are classiﬁed into 15 groups considering symmetry conditions as indicated in Fig. 5.21. The variables Ji (i = 1, . . . , 15) are selected from the predeﬁned lists of available sections. The lists for the columns and beams are shown in Table A.2 and Table A.3 in Appendix A.8, respectively. Note that the number ti of the available sections is 8 for columns and

Optimization of Building Frames

247

TABLE 5.9:

Comparison of performances of frame optimization with various numbers of analyses. Strategy 1, 5000 Minimum Maximum Mean Standard deviation Strategy 1, 200 Minimum Maximum Mean Standard deviation Strategy 2, 200 Minimum Maximum Mean Standard deviation Strategy 3, minimum F1 , 200 Minimum Maximum Mean Standard deviation Strategy 3, minimum F2 , 200 Minimum Maximum Mean Standard deviation Strategy 4, 200 Minimum Maximum Mean Standard deviation

GD (×10−3 )

SP (×10−2 )

SN

1.349 2.784 1.794 0.314

0.968 3.610 1.576 0.516

1.190 2.299 1.815 0.207

9.215 25.52 14.31 0.351

1.112 3.674 2.198 0.671

0.584 1.463 1.033 0.208

7.765 33.17 12.46 4.342

0.923 4.931 2.448 0.897

0.555 1.314 0.988 0.180

7.733 30.16 15.82 5.495

0.922 4.865 2.501 0.823

0.721 1.467 1.072 0.191

7.341 36.88 14.32 5.320

1.188 6.248 2.509 0.998

0.633 1.275 0.955 0.141

7.140 27.81 13.10 4.074

1.186 6.658 2.261 1.008

0.859 1.650 1.207 0.222

9 for beams. The range parameters of the objective values in (5.90) are δ1 = 4.0 and δ2 = 30.0. A uniform random number 0 ≤ r < 1 is generated, and the section is increased or decreased by 1 for r ≥ 0.5 or r < 0.5, respectively. If the value of Ji generated by the random number turns out to be less than 1 or greater than ti , then Ji is replaced by 1 or ti , respectively. All variables are modiﬁed simultaneously when generating a neighborhood solution. Thirty sets of Pareto solutions are generated from diﬀerent initial random seeds. The solid line in each of the following ﬁgures shows the Pareto front obtained by carrying out a local search 90 times with 5000 analyses, which is conceived as a good approximate Pareto front. The optimization results

248

Optimization of Finite Dimensional Structures 60

Compliance

55 50 45 Seed solution 40 35 30 4.5

5

5.5

6

6.5

7

7.5

8

8.5

Total structural volume

FIGURE 5.24: Distribution of the neighborhood solution of the plane frame.

with various numbers of analyses are listed in Table 5.9, where the minimum, maximum, mean, and standard deviation are computed from 30 trials from diﬀerent initial random numbers. Note that ER is not computed, because the accurate Pareto set is not available. The number q of neighborhood solutions is 10 for 5000 analyses and 4 for 200 analyses. It is conﬁrmed from Table 5.9 that a smaller number of analyses leads to larger GD and SP as well as smaller SN. The optimization results are presented by using an abbreviation, e.g., Case 200-2-SN and Case 400-3-F2-GD, in the same manner as the examples of the mathematical problem. The ‘+’ marks in Fig. 5.24 show the 100 neighborhood solutions in the objective function space of a seed solution indicated by the ﬁlled square. In this example, there is no neighborhood solution that improves both of the two objective functions. Therefore, it seems that approaching the Pareto front is more diﬃcult for this frame optimization problem than for the mathematical problem. The result of Case 5000-1-SN is shown in Fig. 5.25(a), which shows that the Pareto set of good accuracy and diversity has been found. In contrast, no solution was found in the correct set of Pareto solutions for Case 200-1GD, as shown in Fig. 5.25(b). If we decrease q to 2 to increase the number of selection steps of the seed solution for Case 200-1-GD, it turns out that there is no candidate solution before reaching the ﬁnal 200th step. The result of Case 200-2-SN is shown in Fig. 5.25(c). In this case, the use of Strategy 2 does not have a signiﬁcant eﬀect, as expected from the distribution of the neighborhood solutions in Fig. 5.24. Fig. 5.25(d) shows the Pareto set for Case 200-3-F1SN, which veriﬁes that the solutions with small F1 are successfully found. In contrast, if we minimize F2 , then the solutions with small F2 are found as shown in Fig. 5.25(e). Finally, the result of Case 200-4-SN using the sharing

70

70

60

60

Compliance

Compliance

Optimization of Building Frames

50

249

50

40

40

30

30 4

5

6

7

8

9

4

10

5

6

7

8

9

10

9

10

9

10

Total structural volume

Total structural volume

(b) Case 200-1-GD

(a) Case 5000-1-SN 70

70

60

60

Compliance

Compliance

A

50

50

40

40

30

30 4

5

6

7

8

9

4

10

5

Total structural volume

(c) Case 200-2-SN

7

8

(d) Case 200-3-F1-SN

70

70

60

60

Compliance

Compliance

6

Total structural volume

50

C

40

50

B 40

30

30 4

5

6

7

8

9

10

4

5

Total structural volume

(e) Case 200-3-F2-SN FIGURE 5.25:

6

7

8

Total structural volume

(f) Case 200-4-SN

Pareto solutions of a plane frame.

function with γ = 0.1 is shown in Fig. 5.25(f). As is seen, Pareto solutions with good diversity can be found using the sharing function, although accuracy is sacriﬁced. The average number of approximate Pareto solutions for the 30 trials is 26.8. Therefore, a good number of solutions are found within 200 analyses. The frames corresponding to the solutions A, B, and C indicated in the objective function space in Figs. 5.25(d)–(f) are shown in Fig. 5.26, where the width of each member is proportional to its cross-sectional area. It is seen from these ﬁgures that the cross-sectional areas are almost uniform if

250

Optimization of Finite Dimensional Structures

(a) Pareto solution A FIGURE 5.26:

(b) Pareto solution B

(c) Pareto solution C

Cross-sectional areas of Pareto optimal solutions.

the total structural volume is small, and an increase of the total structural volume in the Pareto set leads to an increase of cross-sectional areas from the lower stories. This way, the designer can select the most preferred design from the set of Pareto solutions in view of other structural and nonstructural performance measures.

5.5 5.5.1

Multiobjective seismic design of building frames Introduction

Building frames have been conventionally designed in the framework of speciﬁcation-based design, where the bounds on geometries and material properties are given in design codes for the members and connections. However, since the 1990s, performance-based design (PBD) has been widely accepted as a structural design philosophy in which the design criteria are expressed in terms of achieving explicit performance requirements when the structure is subjected to various speciﬁed levels of seismic motions (Mahin, Malley, and Hamburger 2002). The structural performance of an ordinary building frame is usually deﬁned as: (1) resist an occasional strong earthquake without structural damage, (2) allow repairable structural damage against a rare major earthquake, and (3) resist a maximum credible earthquake without total collapse. The performance measures may include the response stresses, maximum load-carrying capacity, interstory drifts, responses and energy dissipation at the limit state or the target (demand) damage state, etc. In the simple procedure of allowable stress design, the stresses under design loads should be within the speciﬁed bound, and the ductility ratio obtained by pushover analysis under monotonic loading is often used as the performance measure. However, for PBD, it is not reasonable to deﬁne performance by the ductility ratio of the story, because yielding may not lead to a total collapse, and a frame can resist seismic excitation by energy dissipation under cyclic plastic deformation (Leelataviwat, Goel, and Stojadinovi´c 2002). Therefore, performances against

Optimization of Building Frames

251

seismic motions are deﬁned in view of displacements or global deformation under speciﬁed seismic motions (Whittaker, Constantinou, and Tsopelas 1998). Xue and Chen (2003) presented a displacement-based PBD using a capacityspectrum method. Hasan, Xu, and Grierson (2002) presented a method of pushover analysis for measuring the plasticity factors at the demand deformation level. However, time-history analysis is so far believed to be the most accurate methodology for evaluating structural performance in the inelastic range against seismic motions. Optimization methods can be eﬀectively used for PBD, because structural performances can be naturally incorporated as objective functions or constraints into the optimization problem (Foley 2002). Bhatti and Pister (1981) may be cited as the ﬁrst paper that considered responses for two levels of seismic motions in the optimization problem. Ganzerli, Pantelides, and Reaveley (2000) utilized the convex model to account for uncertainty of the input motions. Since structural responses should be evaluated many times in the optimization process, the seismic responses may be approximately evaluated by a response spectrum approach. Mohammadi, El Naggar, and Moghaddam (2004) presented an equivalent linearization approach with an inelastic response spectrum for PBD. The capacity spectrum approach is also commonly used, because the fundamental eigenmode dominates in the response of a building frame (Freeman 2004). However, rapid development in computer technology has enabled us to use time-history analysis for a mathematical programming approach to structural optimization (Balling, Pister, and Ciampi 1983). As mentioned earlier, several performances should be simultaneously considered in PBD. Therefore, the optimization problem should be formulated as a multiobjective programming (MOP) problem. Liu, Wen, and Burns (2005) presented a multiobjective optimization method for PBD of frames. Xu, Gong, and Grierson (2006) presented a seismic multiobjective optimization method considering uniform interstory drift under equivalent static loads as one of the objective functions. In MOP for PBD, some of the performance measures are constrained strictly as hard constraints based on a building code, i.e., those performance measures should not exceed the speciﬁed upper bounds regardless of the structural cost. On the other hand, some responses and performance measures called soft targets are to be minimized or maximized only when other targets are not too much sacriﬁced. Note that a soft target is equivalent to the ﬂexible constraint for the constraint satisfaction problem within the framework of decision making in the fuzzy set theory (Dubois and Fortemps 1999). In this section, an optimization method based on time-history analysis and MOP is presented for PBD of steel frames (Pan, Ohsaki, and Kinosh*ta 2007).

252

5.5.2

Optimization of Finite Dimensional Structures

Formulation of the multiobjective programming problem

Optimal designs are found for plane frames subjected to recorded ground motions. The following performance measures may be considered: 1. Total structural volume that is regarded as representing the structural material cost; a more rigorous deﬁnition of the cost, which was described in Sec. 5.1.5, may be used, if possible. 2. Maximum interstory drift angles under seismic motions scaled to Levels 1, 2, and 3, corresponding to occasional, rare, and maximum credible seismic motions. 3. Maximum ﬂoor accelerations under seismic motions scaled to Levels 1, 2, and 3. Since the upper bounds for the maximum interstory drifts are given strictly by the design code, they are regarded as hard constraints and are assigned as inequality constraints in the optimization problem. In contrast, the total structural volume and the maximum ﬂoor accelerations should preferably be minimized, and their bounds are not given explicitly. Hence, they are regarded as soft targets that are included as objective functions of an MOP problem. The beams and columns are classiﬁed into groups, and the members in each group have the same cross-sectional properties. Let Ai denote the crosssectional area of the members in the ith group. From the available standard sections of the square tube columns and wide ﬂange beams, the second moment of inertia Ii and the plastic modulus Zip for the members in the ith group are deﬁned as functions of Ai , as presented in Appendix A.8. Let Li denote the sum of the lengths of members in the ith group. The total structural volume V (A) is given as a function of the design variable vector A = (A1 , . . . , Am )⊤ as m ∑ V (A) = Ai Li (5.101) i=1

where m is the number of groups. The maximum interstory drift angle of the jth story and the maximum acceleration of the jth ﬂoor against the Level (k) (k) k input motion are denoted, respectively, by dj (A) and aj (A), which are (k)

(k)

functions of A. Note that dj (A) and aj (A) are deﬁned as the largest values, respectively, among the earthquake motions of the speciﬁed level if (k) multiple motions are considered. The maximum ﬂoor acceleration amax (A) among all ﬂoors for the Level k input is given as a(k) max (A) =

max

j=1,...,nF

(k)

aj (A)

(5.102) (k)

where nF is the number of stories (ﬂoors). The upper bound for dj (A) is denoted by d(k)U . The multiobjective optimization problem considering input

Optimization of Building Frames

253

motions of three levels is formulated as

where

5.5.3

Minimize

(2) (3) V (A), a(1) max (A), amax (A), and amax (A)

(5.103a)

subject to

(k) dj (A) ≤ d(k)U , ALi ≤ Ai ≤ AU i ,

(5.103b)

ALi

and

AU i

F

(j = 1, . . . , n ; k = 1, 2, 3) (i = 1, . . . , m)

(5.103c)

are the lower and upper bounds for Ai , respectively.

Optimization method

Responses to the recorded motions are computed by time-history analysis considering geometrical and material nonlinearities. Note that the sensitivity coeﬃcients of maximum responses are discontinuous with respect to the design variables, because the occurrence time of the maximum response varies discontinuously due to design modiﬁcation. It is possible to give many constraints at discretized time steps to avoid diﬃculties due to discontinuity (Hsieh and Arora 1984; Paeng and Arora 1989; Reemtsen and R¨ uckmann 1998). For (k) example, the constraint on dj (A) can be replaced as (k)

dj (A, ti ) ≤ d(k)U , (i = 1, . . . , nt )

(5.104)

t

where ti is the ith speciﬁed time step, and n is the number of time steps at which the constraints are given. However, the computational cost for this formulation is very large, because we should have many constraints for an accurate estimation of the maximum acceleration. Therefore, the formulation of Problem (5.103) is used, and a ﬁnite diﬀerence approach is utilized for the sensitivity analysis of maximum responses in the numerical examples. Note that nonlinear programming (NLP) with line search can converge to a local optimum even though the sensitivity coeﬃcients are discontinuous. There have been many studies on sensitivity analysis of elastoplastic responses (Ohsaki and Arora 1994), including dynamic responses (Kleiber 1997). However, these methods are computationally expensive for optimization of complex structures. Furthermore, the optimization problem considered here is highly nonlinear, and there may exist many local optimal solutions. However, an approximate solution can be successfully obtained using an NLP algorithm starting with several initial solutions and selecting the best solutions among the converged solutions (Kim, Haftka, Mason, Watson, and Grossman 2002). A constraint approach is used for solving MOP problem; see Appendix A.4 (k) for details. If minimization of V (A) is more important than that of amax (A), the following single objective problem (SOP) is ﬁrst solved: Minimize

V (A)

subject to

(k) dj (A) ≤ d(k)U , ALi ≤ Ai ≤ AU i ,

(5.105a) F

(j = 1, . . . , n ; k = 1, 2, 3) (i = 1, . . . , m)

(5.105b) (5.105c)

Optimization of Finite Dimensional Structures 11 9 9 9

5 5 5

8 6

7

8

6

7

1

2 12.00

5

7

6

4

5

7

6

8

6

3

5

7 6

5

3

2

5

[emailprotected]

Unit: m 7

1

11

10

11 7

12.00

12

11

10

11

9 9

8

10

7

9

6

8

6

7

5

7

5

3

5

3

1

7

6

7

6

3

2

3

8 4 4

2 12.00

6 6 2 2

6.00

(a) R4

FIGURE 5.27:

12

11 10

10

1

1

6.00

5

12 10

[emailprotected]

254

1 12.00

(b) R8

Three-span plane frame models R4 and R8.

Let Ve denote the optimal objective value of Problem (5.105). A small relaxation parameter for V (A) is denoted by ∆V . We may increase V (A) to (r) Ve + ∆V if one of amax (A) (r ∈ {1, 2, 3}) can be drastically reduced. Suppose (r) (s) minimizing amax (A) is more important than minimizing amax (A) (r ̸= s). Then, the following SOP is next solved: Minimize

a(s) max (A)

subject to

(k) dj (A)

(5.106a)

≤d

(k)U

F

, (j = 1, . . . , n ; k = 1, 2, 3)

V (A) ≤ Ve + ∆V ALi

≤ Ai ≤

AU i ,

(5.106b) (5.106c)

(i = 1, . . . , m)

(5.106d) (t)

Problem (5.106) can be solved consecutively; i.e., we next minimize amax (A) (s) (t ̸= s) under constraint on the relaxed upper bound of amax (A).

5.5.4 5.5.4.1

Numerical examples Descriptions of plane frame models

Optimal designs are found for two three-span plane steel frames, as shown in Figs. 5.27(a) and (b), designated R4 and R8, respectively. Weights of 60 kN/m are distributed on the ﬂoors. Considering the symmetry and practical requirements in the construction process, the columns and beams are classiﬁed into 8 and 12 groups for R4 and R8, respectively, as shown in Fig. 5.27, where the members in each group have the same cross-sectional properties. The cross-sections of columns and beams are taken to be square tubes and wide ﬂanges, respectively. The second moment of inertia Ii and the plastic modulus

Optimization of Building Frames

255

Acceleration response

30 25 20 15 10 5 0 0

1

2

3

4

Natural period

FIGURE 5.28: Pseudoacceleration response spectra for 5% damping of the adopted ground motions scaled to Level 2; solid line: El Centro NS, dashed line: Hachinohe NS, dotted line: Taft EW.

Zip are assumed to be functions of Ai , as shown in Appendix A.8. The upper and lower bounds for Ai are 0.1600 m2 and 0.0060 m2 , respectively, for all the groups of columns and beams. Three recorded near-fault ground motions are adopted, namely, the NS component of 1940 El Centro, the NS component of 1968 Hachinohe, and the EW component of 1952 Taft. Three performance levels, designated as Levels 1, 2, and 3, commonly adopted in Japanese design practice are considered. The peak ground velocities (PGVs) are speciﬁed as 0.25 m/s, 0.50 m/s, and (k) (k) 0.75 m/s, respectively, for Levels 1, 2, and 3. Hence, dj (A) and aj (A) for each seismic level are the maximum values among the three recorded motions. The pseudoacceleration response spectra of the adopted ground motions scaled to Level 2 are plotted in Fig. 5.28 for 5% damping. The upper bounds d(k)U of the maximum interstory drift angles are 0.005, 0.01, and 0.02 for Levels 1, 2, and 3, respectively, for all stories. A structural analysis program called CLAP (Ogawa and Tada 1994) is used for nonlinear dynamic response analysis. This program adopts the concentrated plastic hinges assigned at member ends, where the interaction between axial force and bending moment is considered. Gravity loads are applied prior to the dynamic analysis, and its inﬂuence on plastiﬁcation and the P −∆ eﬀect is included. The modiﬁed method of feasible directions in library DOT Ver. 5.0 (VR&D 1999) is used for optimization, and the forward ﬁnite diﬀerence approach is used for design sensitivity analysis. Because the problem considered here has many local optimal solutions, optimization is carried out starting with 15 diﬀerent initial solutions, where uniform random numbers are generated to assign the initial values of the design variables within the feasible regions.

256

Optimization of Finite Dimensional Structures Volume: 3.99 m3

Volume: 4.24 m3

Volume: 3.64 m3

(a) Cases 1, 2

(b) Case 3

(c) Cases 4, 7, 8, 11, 14

Volume: 4.46 m3

Volume: 3.75 m3

Volume: 6.82 m3

(d) Case 5

(e) Cases 6, 9

(f) Case 10

Volume: 5.76 m3

Volume: 3.81 m3

Volume: 4.54 m3

(g) Case 12

(h) Case 13

(i) Case 15

FIGURE 5.29: Single-objective optimal designs of R4 for Level 1 motions starting with 15 diﬀerent initial solutions.

5.5.4.2

Single-objective optimization

We ﬁrst solve the single-objective optimization problem for minimizing the total structural volume under constraints on interstory drift angles against Level 1 input only in order to verify the convergence property of the optimization algorithm. The optimal designs of R4 are plotted in Figs. 5.29(a)–(i) for the 15 diﬀerent initial solutions indicated by Cases 1–15. The width of each member in the ﬁgures is proportional to its cross-sectional area. As is seen, nine diﬀerent local optimal solutions have been obtained. The optimal objective values are also shown in Figs. 5.29(a)–(i). Cases 4, 7, 8, 11, and 14 converge to the identical optimal design that is also the best solution. Hence, this solution is taken as the approximate global optimal solution. In each of the following optimization problems, the best solution is ﬁrst selected from the local solutions with 15 randomly selected initial solutions. Optimal solutions for ground motions of Levels 1, 2, and 3 separately are as shown in Figs. 5.30(a), (b), and (c), respectively. The total structural volume of R4 is 3.64 m3 for Level 1, which is 1.28 and 1.62 times as large as those for Levels 2 and 3, respectively. Similarly, the total structural volume of R8 is 9.01 m3 for Level 1, which is 1.55 and 1.87 times as large as those for Levels 2 and 3, respectively. Note that both of the PGV and allowable drift angle for Level 2 are twice as large as those for Level 1; however, the total

Optimization of Building Frames

257

Volume: 9.01 m3

Volume: 3.64 m3

(a) Level 1 Volume: 5.83 m3

Volume: 2.85 m3

(b) Level 2 Volume: 4.83 m3

Volume: 2.25 m3

(c) Level 3 FIGURE 5.30: Single-objective optimal designs of R4 and R8 for Levels 1, 2, and 3.

structural volume of Level 2 is smaller than that of Level 1 owing to plastic energy dissipation. The interior columns and beams have larger cross-sections than the exterior members for both structures for all input levels. This is mainly because increasing the cross-sectional areas of the beams and columns at shorter spans is more eﬀective for achieving large lateral stiﬀness than increasing those of longer spans. However, such an eﬀect is less signiﬁcant for Levels 2 and 3 than for Level 1. The cross-sections of lower exterior columns have large values for Level 3, because the lowest exterior columns have the largest axial force and bending moment due to lateral dynamic loads as well as the P − ∆ eﬀect. The members in the lower stories have larger cross-sections than those in

258

Optimization of Finite Dimensional Structures 2

Maximum acceleration amax (m/s ) and ﬁrst natural period T (s) of the multiobjective optimal designs of R4 and R8 with relaxed upper-bound total structural volume. R4 R4 amax T amax T Single objective 10.68 0.56 7.05 0.98 Relax 10% 9.96 0.53 5.02 1.26 Relax 20% 9.50 0.48 4.80 1.27

TABLE 5.10:

the higher stories; i.e., no additional constraint is needed so that the column and beam have smaller cross-sections, respectively, than those directly below them, because the upper-bound drift angles d(k)U are the same for all stories, whereas the shear and axial forces in the lower stories are larger than those in the higher stories. We can see from Fig. 5.30 that the columns generally have larger cross-sections than the beams, which conforms with the design philosophy of strong column-weak beam. Optimization is next carried out considering Levels 1, 2, and 3 inputs simultaneously. The optimal objective values for R4 and R8 are 3.65 m3 and 9.04 m3 , respectively. The distribution of cross-sectional areas is similar to Fig. 5.30(a) considering Level 1 only, because constraints on the interstory drift angle (0.005) of Level 1 are signiﬁcantly stricter than those speciﬁed in Levels 2 and 3. 5.5.4.3

Multiobjective optimization

The constraint approach to MOP is applied considering Level 1 input only. We investigate the results of relaxation of the total structural volume by 10% and 20%; i.e., ∆V in Problem (5.106) is 0.1Ve and 0.2Ve , respectively. 2 The maximum ﬂoor accelerations amax (m/s ) are listed in Table 5.10. As is seen, the maximum ﬂoor accelerations are reduced by 6.7% and 11%, while sacriﬁcing 10% and 20% of the structural volume, respectively, for R4, whereas they are reduced by 30% and 32% for R8. The values of T (s) are 0.56 and 0.98 for the original designs of R4 and R8, respectively. They are shifted to 0.53 and 1.26 by relaxing 10% of the structural volume for R4 and R8, respectively, and to 0.48 and 1.27 by relaxing 20% of the structural volume. The maximum ﬂoor acceleration is reduced mainly due to the shift of the natural period. As is seen from the response spectra in Fig. 5.28, the response acceleration is a roughly decreasing function of T in the region T > 1.0; hence, a large reduction in acceleration is observed for R8 as a result of the increase of T from 0.98 to 1.26.

Chapter 6 Optimization of Spatial Trusses and Frames

In this chapter, we present various optimization results of spatial frames, namely, latticed domes and long-span arches. Following the historical review in Sec. 6.1, we carry out sensitivity analysis and optimization of archtype trusses and a double-layer cylindrical grid in Sec. 6.2, as illustrative examples. In Sec. 6.3, single-point-search heuristic approaches, e.g., greedy method, simulated annealing, and tabu search, are applied to the optimal design of a spatial frame, and their performances are compared. In Sec. 6.4, an approach is presented for incorporating the designer’s preference of shape of an arch-type frame that is described using a B´ezier curve. Multiobjective shape optimization of a single-layer latticed shell is presented in Sec. 6.5. In Sec. 6.6, a method based on the genetic algorithm is presented for conﬁguration optimization of arch-type trusses incorporating explicit geometrical constraints. In Sec. 6.7, a parametric programming approach is presented for estimating the eﬀect of spatial variation of seismic motions on optimal solutions. Finally, a substructure approach is presented in Sec. 6.8 to optimize a roof truss without carrying out analysis of the whole structure at each optimization step.

6.1

Introduction

Spatial trusses and frames, including arches, reticulated shells, and cablesupported frames, are designed and constructed for covering large spaces for stadiums, arenas, etc. For regular building frames, the locations of members are determined by architects mainly in view of planning of the ﬂoors, and the cross-sectional properties of the members are determined by engineers. In contrast, mechanical properties such as displacements, stresses, and buckling loads against speciﬁed design loads play key roles for determination of the shapes and topologies of spatial trusses and frames. Although the preferred shape is drawn by an architect, the beauty of a spatial structure is closely related to the eﬃciency in the load-carrying capacity of the structure. Therefore, structural optimization plays an important role in the design of spatial

259

260

Optimization of Finite Dimensional Structures

structures. Approaches to designing mechanically eﬃcient spatial structures are classiﬁed into form ﬁnding and structural optimization. The former is aimed at designing the shape of ﬂexible structures, including membrane roof structures and cable networks, where computational approaches are developed for obtaining, e.g., the catenary and minimal surface that are optimal for simple loading conditions (Otto 1967; Krishna 1979). Computational approaches were also developed for ﬁnding an ideal shape of equilibrium under complex and realistic boundary conditions (Haber and Abel 1982). In contrast, structural optimization was mainly developed in the ﬁelds of mechanical engineering and aeronautical engineering (Shield 1960; Haug and Cea 1981; Bennett and Botkin 1986; Guillet, No¨el, and L´eon 1996). In this book, we focus on structural optimization methods for application to structures in civil and architectural engineering. It should be noted here that form ﬁnding and structural optimization were recently combined to develop a ﬁeld called structural morphology. Readers can consult many books and papers published by researchers in the society of spatial structures, namely, the International Association of Shell and Spatial Structures, e.g., Pugnale and Sassone (2007), Kimura and Ohmori (2008). Because spatial trusses and frames have many nodes and members, their optimization methods have been developed in accordance with the development of computer technologies. Since the 1970s, many papers have been published on the optimization of spatial trusses; e.g., Pedersen (1973), Saka and Ulker (1991), Krishnamoorthy, Venkatesh, and Sudarshan (2002), Saka (2007). Recently, some evolutionary strategies have been applied to ﬁnd optimal cross-sectional areas, topology, and geometry of spatial frames (Ebenau, Rottsch¨ afer, and Thierauf 2005; Rajasekaran, Mohan, and Khamis 2004). Kaveh, Azar, and Talatahari (2008) applied a heuristic approach called ant colony optimization to stiﬀness design of spatial trusses under stress constraints. Lemonge, Barbosa, and Fonseca (2009) optimized shape and crosssectional areas of latticed domes using a genetic algorithm with constraints on the number of diﬀerent cross-sections to reduce construction cost. In the design process of spatial frames, similar to regular building frames, cross-sectional properties are often selected from the list or catalog of standard sections. Therefore, the optimization problem is formulated as a combinatorial problem, for which the heuristic approaches presented in Secs. 1.12, 5.4, and Appendix A.3 can be eﬀectively used for optimizing real-world structures. Since the designer’s preference plays a key role in the shape design of a spatial structure, its design problem can be formulated as a multiobjective optimization problem considering the geometrical and mechanical properties as the objective functions. Ohsaki, Nakamura, and Kohiyama (1997) presented an optimization approach to the design of a double-layer dome truss considering the properties of the surface and the curves formed by the members. Ohsaki and Hayashi (2000) developed an approach to shape optimization of ribbed shells considering the fairness metrics of the surface (Roulier and

Optimization of Spatial Trusses and Frames

261

Rando 1994). Their method was extended to a multiobjective programming approach for roundness of shape and mechanical eﬃciency of the structure (Ohsaki, Ogawa, and Tateishi 2003; Fujita and Ohsaki 2009). Another unique aspect of designing spatial frames is that the safety against buckling often turns out to be the most critical design requirement. However, optimization against buckling is out of scope of this book, because it was extensively reviewed in the previous book by the author (Ohsaki and Ikeda 2007).

6.2 6.2.1

Seismic optimization of spatial trusses Introduction

The concept called performance-based design has been proposed in countries prone to seismic risk to ensure servicability of structures and life safety, for speciﬁed load levels (Mahin, Malley, and Hamburger 2002). Among the many types of structures, spatial structures are very important as the evacuation facilities at an event of seismic disaster. The key aspects of seismic design of spatial structures are summarized as follows: 1. Vertical motions are not negligible in comparison to horizontal motions for evaluation of seismic responses. Furthermore, vertical responses due to horizontal motions should be considered for a curved roof structure, as shown in Sec. 6.2.3. The constraints should be given for the vertical accelerations of the roof so that the hanging nonstructural components and devices, including ceiling and lights, are not damaged during a severe earthquake. 2. Several modes may dominate in the seismic response, in contrast to a building frame in which only the lowest mode dominates. It is important to incorporate the eﬀect of higher modes especially in acceleration responses. 3. The eﬀects of spatial variation of seismic motions should be considered for long-span structures, e.g., bridges spanning between supports in different soil conditions. In the following, illustrative examples are presented for sensitivity analysis of eigenvalues and eigenmodes as well as optimization of spatial trusses under constraints on seismic responses.

262

Optimization of Finite Dimensional Structures

40 m

FIGURE 6.1:

A 39-bar arch-type truss.

(a) First mode FIGURE 6.2:

6.2.2

(b) Second mode

Eigenmodes of the 39-bar arch-type truss with ϕ = 40 deg.

Design sensitivity analysis

Let K(A) and M(A) denote the n × n stiﬀness matrix and mass matrix, respectively, of a truss, which are functions of the vector A = (A1 , . . . , Am )⊤ of the cross-sectional areas of m members. The mass matrix consists of the structural (member) mass and the nonstructural (nodal) mass. The rth eigenvalue and eigenmode are denoted by Ωr and Φr , respectively. The sensitivity coeﬃcients of Ωr with respect to Ai are given as (see Sec. 2.3) ∂Ωr = Φ⊤ r ∂Ai

(

∂K ∂M − Ωr ∂Ai ∂Ai

) Φr

(6.1)

Because K and M of a truss are linear functions of Ai , the matrices ∂K/∂Ai and ∂M/∂Ai are constant. Furthermore, if the structural mass is negligibly small compared with the nonstructural mass, then Ωr is generally a nondecreasing function of Ai , because ∂K/∂Ai is positive semideﬁnite, except for the case where there exists a member that has a nonzero mass density and no deformation in the eigenmode. The sensitivity coeﬃcient of the rth natural frequency fr is computed from those of the eigenvalues as ∂fr 1 ∂Ωr = ∂Ai 8π 2 fr ∂Ai

(6.2)

Example 6.1 Consider, as an illustrative example, a 39-bar arch-type truss, as shown in Fig. 6.1, where the span length is 40 m, the open angle ϕ of the lower circle is 40 deg, and the diﬀerence between the radii of the lower and upper circles is 2 m. The lengths of the lower chords, upper chords, and diagonals are the same, respectively. The material of the member is steel, where the mass 3 2 density is 7.86×10−6 kg/mm and the elastic modulus is 210.0 kN/mm . The 5 concentrated mass of 1.0 × 10 kg exists at each node.

Optimization of Spatial Trusses and Frames

FIGURE 6.3:

263

Sensitivity coeﬃcients of f1 of the 39-bar arch-type truss.

If the cross-sectional areas are 0.01 m2 for all members, the two lowest frequencies (Hz) are f1 = 2.0621 and f2 = 2.2258, and the corresponding eigenmodes are as shown in the solid lines in Fig. 6.2, while the dotted lines indicate the undeformed shape. Because the ratio of the rise to the span of the arch is relatively small, the ﬁrst mode is symmetric with respect to the vertical center axis, while the second mode is antisymmetric. Note that the second mode is excited by a horizontal motion; therefore, as is seen from the mode shape in Fig. 6.2(b), the vertical response is induced by the horizontal input at the nodes except the lower node on the center line. The sensitivity coeﬃcients ∂f1 /∂Ai of f1 with respect to the cross-sectional areas are plotted in Fig. 6.3, where the width of each member is proportional to the sensitivity coeﬃcient, which is positive for all members; i.e., an increase of the cross-sectional area of any member results in an increase of f1 . For example, the sensitivity coeﬃcient with respect to the cross-sectional area of member 1, which is the lower chord connected to the support, is 1.2364 × 10−5 mm−2 ; i.e., an increase of 1 mm2 of A1 leads to an increase of f1 of 1.2364 × 10−5 Hz. This way, by carrying out design sensitivity analysis, the designer can obtain valuable information on the eﬀect of design modiﬁcation in view of increasing the lowest natural frequency.

6.2.3

Optimization against seismic excitations

Optimal solutions are found for an arch-type truss and a cylindrical doublelayer grid subjected to seismic excitations. The response spectrum by Newmark and Hall (1982) is used, and the modal responses are combined using the complete quadratic combination (CQC) method (Wilson, Der Kiureghian, and Bayo 1982) (see Appendix A.7 for details of the response spectrum approach). The seismic motions in the horizontal and vertical directions are simultaneously considered, where the level of the vertical motion is 2/3 of those for the horizontal motions; i.e., the values of the parameters for maximum acceleration, velocity, and displacement of the ground motion in (A.154) 2 in Appendix A.7 are CA = 2.01 m/s , CV = 0.25 m/s, and CD = 0.1875 m for 2 horizontal motion, and CA = 1.34 m/s , CV = 0.1667 m/s, and CD = 0.125 m for vertical motion. Responses to vertical and horizontal motions evaluated using the CQC method are further combined using the square-root-of-sumof-square (SRSS) method as presented by Semby and Der Kiureghian (1985), because these two directions coincide with the principal directions of the struc-

264

Optimization of Finite Dimensional Structures

(a) Cross-sectional area FIGURE 6.4:

(b) First eigenmode

Optimal solution for ϕ = 40 deg.

(a) Cross-sectional area

(b) First eigenmode

FIGURE 6.5:

Optimal solution for ϕ = 120 deg.

FIGURE 6.6:

A 7 × 7 double-layer cylindrical grid.

ture. The sensitivity coeﬃcients of the modal responses can also be obtained analytically, as shown in Appendix A.7. The cross-sectional areas are assumed to be continuous variables, and sequential quadratic programming is used for optimization. Optimal solutions are ﬁrst found for the arch-type truss in Fig. 6.1. The 2 upper bound for the absolute value of stress is 105 N/mm , and the lower2 bound cross-sectional area is 500 mm for all members. The optimal solutions for the open angle ϕ = 40 deg and 120 deg are shown in Figs. 6.4(a) and 6.5(a), respectively, where the width of each member is proportional to its cross-sectional area, and the ﬁrst eigenmodes are shown in Figs. 6.4(b) and 6.5(b), respectively. As is seen, the optimal truss for ϕ = 40 deg has an antisymmetric ﬁrst eigenmode, whereas it is symmetric if all the members have the same cross-sectional area, as shown in Fig. 6.2(a). The two lowest frequencies of the optimal truss with ϕ = 40 deg are 2.1397 Hz and 2.6687 Hz, which are moderately close, and correspond to antisymmetric and symmetric modes, respectively. There-

Optimization of Spatial Trusses and Frames

265

longitudinal direction

circular direction

(a) Diagonal view longitudinal direction

(b) Lower chords longitudinal direction

circular direction

circular direction

(d) Diagonals (c) Upper chords FIGURE 6.7: Optimal solution of the 7 × 7 double-layer cylindrical grid.

fore, due to the contribution of the second eigenmode to the responses against vertical motion, the optimal cross-sectional areas have large values around the center, because the second eigenmode has a large curvature change assuming that the arch-type truss represents a continuous arch. On the other hand, the antisymmetric ﬁrst mode dominates in response of the optimal truss with ϕ = 120 deg; hence, the cross-sectional areas have large values around the regions between the center and the supports. We next ﬁnd the optimal solutions for a 7 × 7 double-layer cylindrical grid, as shown in Fig. 6.6, which has pin supports at the four lower corners and lumped mass of 1000 kg at each node. The members are pin-jointed at the nodes. The span length between the supports is 12 m for both the longitudinal and circular directions. The open angle of the lower cylinder is 80 deg, and the distance between the lower and upper cylinders is 2 m. The lower and upper chords in each direction and the diagonals have the same lengths, respectively. The material parameters, the lower-bound cross-sectional areas, and the upper-bound stress are the same as those of the arch-type truss. The truss is subjected to three-directional seismic motions. The optimal cross-sectional areas are shown in Figs. 6.7(a)–(d). As is seen, the upper chords in the longitudinal direction between the supports, and the lower chords and the diagonals near the supports have large cross-sectional areas. The lowest three frequen-

266

Optimization of Finite Dimensional Structures y

x

z x 32 m

FIGURE 6.8:

A single-layer spatial frame.

cies (Hz) are 2.7964, 2.8743, and 3.1496, which are very close. The ﬁrst mode is symmetric, the second mode is antisymmetric with respect to the plane that is perpendicular to the circles, and the third mode is antisymmetric with respect to the plane that is parallel to the circles. These modes are excited by vertical motion and horizontal motions in two directions, respectively.

6.3

Heuristic approaches to optimization of a spatial frame

In the design process of spatial frames, the cross-sectional properties are often selected from the list or catalog of standard sections. Therefore, the optimization problem is formulated as a combinatorial problem, for which the heuristic approaches presented in Secs. 1.12, 5.4, and Appendix A.3 are eﬀectively used for spatial trusses and frames. In this section, we summarize the results by Ohsaki (2005b) and investigate the performance of single-pointsearch heuristics to the design of a spatial frame with discrete cross-sectional properties. Consider a 132-bar single-layer spatial frame, as shown in Fig. 6.8, subjected to static loads. All the nodes are on a sphere with open angle 40 deg and span length 32 m. The members in the longitudinal (meridian) and circumferential directions, respectively, have the same lengths. The members

Optimization of Spatial Trusses and Frames

267

15 14

10 6

13

9 3

12 5

8

1 2

FIGURE 6.9:

4

7

11

Node numbers of the single-layer spatial frame.

TABLE 6.1: spatial frame. Node number 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15

Nodal coordinates (mm) of the single-layer x

y

0.0 4077.22 2038.60 8123.42 7035.08 4061.70 12,107.78 11,377.60 9275.10 6053.88 16000.0 15,454.82 13,856.40 11,313.70 8000.0

0.0 0.0 3530.98 0.0 4061.70 7035.08 0.0 4141.10 7782.74 10,485.64 0.0 4141.10 8000.0 11,313.70 13,826.40

z 2821.24 2643.22 2643.22 2110.52 2110.52 2110.52 1227.22 1227.22 1227.22 1227.22 0.0 0.0 0.0 0.0 0.0

are rigidly jointed at the nodes and pin-jointed at the supports along the lowest circle. The node numbers and nodal coordinates of one of the six equal parts are shown in Fig. 6.9 and Table 6.1, respectively. The elastic modulus 2 2 is 200 kN/mm , and the weight density is 7.7 × 10−5 N/mm . A concentrated load of 40 kN is applied in the negative z-direction at each node, and the self-weight of members is added to the nodal loads. The members have cylindrical cross-sections with the external and internal radii denoted by R and r, respectively. For simplicity, we assume the relation between R and r as r = 0.96R (6.3) Therefore, all the cross-sectional properties, including the cross-sectional area and the second moment of inertia, are functions of R. The members are

268

Optimization of Finite Dimensional Structures 9 10 8

6

10

7 5

3

8

7

4 7 2

1 1

FIGURE 6.10:

TABLE 6.2: Member group 1 2 3 4 5 6 7 8 9 10 V (m3 ) Maximum stress ratio No. of analyses

5 4 3

10 10 10

7

8 10

6

9

Member groups of the single-layer spatial frame. Optimal solutions using various approaches (Part 1). Continuous Greedy Enumeration Greedy Stingy (Type 1) (Type 2) 184.2 4 4 4 4 60.0 1 1 1 1 156.3 3 3 4 3 116.2 2 2 1 2 109.2 2 2 3 3 153.0 3 3 4 4 126.9 3 3 3 3 117.9 2 2 1 1 60.0 2 1 2 2 144.5 3 3 3 3 2.504 3.831 3.766 3.996 4.005 1.000 0.9957 0.9746 0.9409 0.8961 1233

1237

489, 331

16

14

classiﬁed into 10 groups, as shown in Fig. 6.10, and the external radius of the members in group i is denoted by Ri . A list R is given as follows, from which Ri of each group is to be selected: R = {60, 120, 180, 240}

(6.4)

where the unit of length is mm. The objective function to be minimized is the total structural volume V . Let σ A denote the stress at a member end due to axial force. The maximum stresses at the edge of a member end due to bending around the mutually perpendicular two axes are denoted by σ1B and σ2B . Because each member has a circular cylindrical section, the maximum absolute value σ max of the stress

Optimization of Spatial Trusses and Frames is estimated by σ max = |σ A | +

269

√ (σ1B )2 + (σ2B )2

(6.5)

2

The upper bound 50 N/mm is given for σ max of each member. In the following, the ratio of σ max to the upper-bound stress is called the stress ratio. An optimal solution is ﬁrst found by considering Ri as a continuous variable. Optimization is carried out using IDESIGN Ver. 3.5 (Arora and Tseng 1987), where sequential quadratic programming is used. The lower and upper bounds for Ri are 60 mm and 240 mm, respectively, which are equal to the smallest and largest discrete values in the list R. The optimal solution is listed in the second column of Table 6.2. Note that the number of analyses before reaching the optimal solution is 1233, which is very large, because a ﬁnite diﬀerence approach is used for design sensitivity analysis. The maximum value of the stress ratio of the members is 1.0 for all groups, which means that the optimal solution is fully stressed. The nearest value of Ri from the continuous optimal solution is selected from the list R for each group; i.e., R1 , . . . , R10 are 180, 60, 180, 120, 120, 160, 120, 120, 60, 160. Since the solution obtained this way does not satisfy the stress constraints, the value of Ri corresponding to the maximum stress ratio among all groups is increased consecutively until stress constraints are satisﬁed in all members. This method is denoted as the greedy method (Type 1), and its results are listed in the third column of Table 6.2, where the numbers 1, 2, 3, and 4 correspond to Ri = 60, 120, 180, and 240, respectively. The number of analyses is 4 for the iterative correction using the greedy method to which is added 1233 for sequential quadratic programming to result in the total number of analyses 1233 + 4 = 1237. To conﬁrm the accuracy of the heuristic methods, the global optimal solution has been found by enumerating all 410 = 1, 048, 574 solutions. The result is listed in the fourth column of Table 6.2. Note that the total number of analyses is less than 410 , because structural analysis is not carried out in the enumeration process if the total structural volume is less than that of the current upper-bound solution satisfying stress constraints in all groups. As is seen, the simple greedy method (Type 1) can reach a good approximate solution. Next we consider a greedy method (Type 2) starting with Ri = 60 mm, which is the smallest value in the list R, for all groups. The result is listed in the ﬁfth column of Table 6.2. On the other hand, if we use the stingy method starting with the initial solution Ri = 240 mm for all groups and reducing the cross-section consecutively, the result is as shown in the sixth column of Table 6.2. In this case, the stingy method reaches a better solution than the greedy method (Type 2). The number of analyses is very small for both methods. The result of simulated annealing (SA) is listed in the second column of Table 6.3, where the quadratic exterior penalty function approach described

270

TABLE 6.3: Member group 1 2 3 4 5 6 7 8 9 10 V (m3 ) Maximum stress ratio No. of analyses

Optimization of Finite Dimensional Structures Optimal solutions using various approaches (Part 2). SA Random TS Enumeration near Greedy search continuous solution (Type 3) 4 4 4 4 4 2 1 1 1 1 3 3 3 3 4 2 2 2 2 2 2 2 2 2 3 3 3 3 3 4 3 3 3 3 3 2 2 2 2 1 1 2 1 1 2 3 3 3 3 3 3.831 3.766 3.766 3.766 4.157 0.9980 0.9746 0.9746 0.8441 0.9746 2000

2000

282

2256

1244

in Appendix A.2.2.4 is used for incorporating the stress constraints, and the penalty parameter is 1011 , which is also used in all the examples below that utilize the penalty function approach. The initial value of the temperature parameter is 1, which is multiplied by 0.99 at each iterative step. We tried 10 diﬀerent random initial solutions with 200 steps for each case. Therefore, the total number of analyses is 2000. The result of a random search is listed in the third column of Table 6.3, where the number of analyses is also 2000. Note that the random search has better performance than SA for this example. The result of a tabu search (TS) is shown in the fourth column of Table 6.3. The maximum number of steps is 50, and the length of the tabu list is also 50, which means that a solution selected once is not chosen again. A maximum of 20 neighborhood solutions are randomly searched at each iterative step by increasing or decreasing the rank of the section by 1 for each variable. As is seen, the global optimal solution has been found by TS with 282 analyses from a randomly generated initial solution. Note that the same optimal solution has been found from ﬁve diﬀerent initial solutions. The ﬁfth column of Table 6.3 shows the result of enumeration in the neighborhood of the optimal solution with a continuous variable (second column of Table 6.2). For example, Ri = 180 and 240 are selected as the candidates for group 1, because R1 = 184.2 for the continuous solution. In this example, there exists a global optimal solution in the neighborhood of the continuous solution. Finally, a greedy method (Type 3) was tried starting with the largest value of Ri in the list that does not exceed the continuous solution in the second column of Table 6.2. The results are listed in the sixth column of Table 6.3. In this case, no good solution was found by this approach.

Optimization of Spatial Trusses and Frames

TABLE 6.4:

Comparison of performances of for various lists of available radii. R∗ 60 58 56 Greedy (Type 1) 3.831 3.831 3.836 Enumeration 3.766 3.682 3.565 Greedy (Type 2) 3.996 4.026 3.754 Stingy 4.005 3.743 3.621 SA 3.831 3.682 3.587 Random search 3.766 3.682 3.586 TS 3.766 3.682 3.565 Enumeration near 3.766 3.728 3.709 continuous solution Greedy (Type 3) 4.157 3.831 3.836

271

the heuristic approaches 54 3.567 3.314 3.709 4.634 3.514 3.314 3.478 3.721

52 3.557 3.259 4.249 4.297 3.259 3.273 3.389 3.450

50 3.186 3.092 4.434 5.092 3.137 3.102 3.092 3.102

3.567

3.557

4.037

The optimization results presented above, however, strongly depend on the geometry of the structure, the available values of Ri , the load level, etc. Therefore, in the following, we parametrically vary the available values of Ri , and compare the performances of the heuristic approaches. Let R∗ denote the unit value of the available cross-sectional areas, and deﬁne the list R as R = {R∗ , 2R∗ , 3R∗ , 4R∗ } (6.6) Note that R∗ = 60 corresponds to the previous example. The optimal objective values for R∗ = 60, 58, 56, 54, 52, and 50 by each method presented above are listed in Table 6.4, where the underline indicates that the value of the optimal solution coincides with that of the global optimal solution obtained by enumeration. It is observed from Table 6.4 that SA, TS, and random search can reach the global solution or a good approximate solution. The greedy methods and the stingy method may sometimes ﬁnd a solution that has a very large objective value. Furthermore, it often happens that the global solution does not exist in the neighborhood of the continuous solution.

6.4 6.4.1

Shape optimization considering the designer’s preference Introduction

In the practical design process of long-span trusses and frames, the stresses and displacements under static and dynamic design loads are required to be within the speciﬁed bounds deﬁned by building codes. Another important aspect in structural design in civil and architectural design is that the most pre-

272

Optimization of Finite Dimensional Structures

(a)

(b) FIGURE 6.11: Optimal shapes of an arch subjected to static loads; (a) self-weight, (b) concentrated loads.

ferred shape may be presented by the designer or architect prior to structural design considering mechanical properties. Therefore, one of the drawbacks in structural optimization under constraints on mechanical properties only is that the solution often turns out to be an unrealistic design that appears to be far from optimum from the geometrical and aesthetic point of view. In this section, we consider the shape optimization problem in which optimal locations of the nodes are to be optimized while topology variation is not allowed (Lin, Che, and Yu 1982; Imai and Schmit 1982; Svanberg 1981; Saka 1980; Sadek 1986) (see Sec. 1.10 for simple examples of shape optimization). Smooth curves and surfaces are described eﬃciently with techniques developed in computer aided geometric design (CAGD) (Faux and Pratt 1979; Farin 1988; Rogers and Adams 1990; Barnhill 1994); see Appendix A.6. A method is presented for the shape design of a plane arch-type frame with nodal locations deﬁned in terms of B´ezier curves (Ohsaki, Nakamura, and Isshiki 1998). The deviation of the curve associated with the lower nodes from the shape preferred by the designer is considered as one of the objective functions. Compliance is also minimized as the performance measure of the stiﬀness. Example 6.2 An arch is traditionally regarded as the ideal shape for carrying vertical loads. However, the optimal shape varies with the distribution of external loads. As an illustrative example, optimal shapes of pin-supported arches are found using a nonlinear programming algorithm. The arch is discretized to 20 beam elements, which have a cylindrical section with the same cross-sectional area and the ﬁxed ratio 0.1 of the thickness to the external radius. Hence, the design variables are the cross-sectional area and the vertical coordinates of the 19 internal nodes between the elements, where the ratio of the rise (height of the center node) to the span is ﬁxed at 0.15. Constraints are given for the maximum edge stresses evaluated at the nodes. The objective function

Optimization of Spatial Trusses and Frames

273

to be minimized is the total structural volume. A detailed description of the model is omitted here, because the purpose of this example is to illustrate the dependence of the optimal shape on the loading conditions. The optimal shape under self-weight is shown in Fig. 6.11(a), which is close to the optimal shape of a hanging cable called a catenary if the vertical coordinates are reversed. Therefore, the arch in Fig. 6.11(a) is called a catenary arch. For the case in which the three concentrated loads are applied around the center, the optimal shape is as shown in Fig. 6.11(b). As is seen, the optimal shape strongly depends on the load distributions, and the trade-oﬀ between the shape preferred by the designer and the mechanically optimal shape should be considered in the practical shape optimization process.

6.4.2

Description of an arch-type frame using a B´ ezier curve

Let P(u) = (x(u), y(u))⊤ denote a B´ezier curve deﬁned by the parameter 0 ≤ u ≤ 1 in the two-dimensional space (x, y). The B´ezier curve of order nC is deﬁned in terms of the control points Ri = (Xi , Yi )⊤ (i = 0, 1, . . . , nC ) as P(u) =

( ) x(u) y(u) n ∑ C

=

C

Ri Bin (u)

(6.7)

i=0

) n ( ∑ Xi C

=

i=0 C

Yi

C

Bin (u)

where Bin (u) is the Bernstein polynomial of order nC (see Appendix A.6 for details of the B´ezier curve). Utilizing the convex hull property, the shape of the B´ezier curve may be indirectly and interactively controlled by modifying the locations of the control points. Another important property of the B´ezier curve is that P(0) = R0 and P(1) = RnC are satisﬁed; i.e., the two ends of the curve coincide with the control points. Furthermore, the curve inscribes the control polygon at R0 and RnC . It is straightforward to optimize the B´ezier curves by considering the locations of the control points as design variables. For the design of a truss or a frame, however, the locations of the nodes need to be deﬁned after the shape of the curve is given. Suppose P(u) in (6.7) deﬁnes the B´ezier curve associated with the lower nodes. A sequence of points on a B´ezier curve is speciﬁed by a set of parameters u = (u1 , . . . , unJ )⊤ , where nJ is the number of nodes of the arch-type frame, including the upper and lower nodes as well as the supports. If the ith node is located on the lower curve, then its coordinate vector is computed as P(ui ).

274

Optimization of Finite Dimensional Structures

dn0 (ui+1)

normal offset vector dn0 (ui+3)

ui+1

ui+2

dnv

ui+1

ui+3

ui lower curve

ui+2

ui+3

ui lower curve (b) Vertical oﬀset vector

(a) Normal oﬀset vector FIGURE 6.12:

vertical offset vector dnv

Process of generating an arch-type frame.

The locations of the upper nodes are deﬁned in terms of the normal oﬀset vector, as illustrated in Fig. 6.12(a), or the vertical oﬀset vector in Fig. 6.12(b) (see Appendix A.6.3 for deﬁnitions of the adjoint curves, including the oﬀset curves). Conversely, the upper nodes can be speciﬁed on a B´ezier curve, and the lower nodes may be deﬁned using the oﬀset vector. Two approaches with the use of normal and constant vertical oﬀset vectors were presented by Ohsaki, Nakamura, and Kohiyama (1997) for double-layer spatial trusses. A vector n(u) that is normal to the lower B´ezier curve P(u) is deﬁned as n(u) = T =T

∂P ∂u nC ∑ i=0

(6.8)

C

Ri

∂Bin (u) ∂u

where T is the rotation matrix of the angle π/2. Then, the unit normal vector n0 is given as n(u) (6.9) n0 (u) = √ n⊤ n The unit vector in the y-direction is deﬁned as nv = (0, 1)⊤ . Let d denote the prescribed length of the oﬀset vector. Then the nodal coordinate vectors are computed as Lower node : P(ui )

(6.10a) 0

Upper node : P(ui ) + dn (ui ) or P(ui ) + dn

v

(6.10b)

The process of generating an arch-type frame is illustrated in Fig. 6.12. Examples of arch-type frames with vertical and normal oﬀset vectors, respectively, as well as control points for the lower curve, are illustrated in Fig. 6.13, where the numbers of lower and upper nodes, except the two supports, are 17 and 18, respectively. Hence, the parameter values for the ith lower and upper nodes from the left are given as i/18 and (2i − 1)/36, respectively.

Optimization of Spatial Trusses and Frames

30 m

30m

(a) Normal oﬀset vector

(b) Vertical oﬀset vector

275

y x

(c) Control points for lower curve FIGURE 6.13: 6.

An arch-type frame modeled using a B´ezier curve of order

Ohsaki, Nakamura, and Isshiki (1998) used a B´ezier function for deﬁning the cross-sectional areas. The B´ezier function A(u) of order nC for the crosssectional area is given as n ∑ C

A(u) =

C

RiA Bin (u)

(6.11)

i=0

where RiA (i = 0, 1, . . . , nC ) are the function values at the control points. Suppose member i connects nodes j and k, and let uA i = (uj + uk )/2. Then, the cross-sectional area of member i is given as Ai = A(uA i ). This way, smoothly distributing cross-sectional areas can be optimized by considering RiA as design variables. Note that diﬀerent B´ezier functions should be used for the lower chords, upper chords, and diagonals. In the following numerical examples, however, we assume that cross-sectional areas are ﬁxed at the speciﬁed values.

6.4.3

Shape optimization incorporating the designer’s preference

A multiobjective programming problem is formulated for obtaining the trade-oﬀ designs between the deviation from the desired shape and the mechanical performance that is deﬁned by compliance under static loads. Let K and U denote the stiﬀness matrix and the displacement vector under speciﬁed loads, which are functions of X consisting of the variable coordinates of the control points. Note that the control points at the supports are usually ﬁxed. Then, compliance W (X) is deﬁned using K(X) and U(X) as W (X) = U(X)⊤ K(X)U(X)

(6.12)

276

Optimization of Finite Dimensional Structures

An upper bound W U is given for the compliance so that the frame has enough stiﬀness against the external loads. Suppose the shape of the lower curve preferred by the designer or architect is expressed using a B´ezier curve as n ∑ C

e P(u) =

e i B nC (u) R i

(6.13)

i=0

In the following, a tilde indicates a speciﬁed value. Using (6.7), the deviation e G(X) of a B´ezier curve P(u) = (x(u), y(u))⊤ from P(u) = (e x(u), ye(u))⊤ is deﬁned as follows: ∫ 1 ⊤ e e G(X) = (P(u) − P(u)) (P(u) − P(u))du ∫

∫

1 0

∫

1

=

(y(u) − ye(u))2 du 0

n ∑ C

2

ei )Bin (u) du (Xi − X C

i=1

∫

1

+

2 C n ∑ C (Yi − Yei )Bin (u) du

(6.14)

i=1

∫ n ∑ n [ ∑ ei )(Xj − X ej ) (Xi − X C

=

1

(x(u) − x e(u))2 du +

=

C

1

C

]

i=1 j=1

∫ n ∑ n [ ∑ e e + (Yi − Yi )(Yj − Yj ) C

C

Bin (u)Bjn (u)du

C

1

C C Bin (u)Bjn (u)du

]

i=1 j=1

It is seen from (6.14) that the integration of the basis functions is to be carried out independently of the design variables X. The multiobjective optimization problem is formulated as Minimize

G(X) and W (X)

subject to X ≤ X ≤ X L

U

(6.15a) (6.15b)

where XL and XU are the lower and upper bounds for X, respectively. A set of Pareto optimal solutions is found using the constraint approach (see Appendix A.4 for details). First, we minimize the shape deviation G(X) and the compliance W (X) independently of side constraints (6.15b) on X by solving the single-objective optimization problem. Then appropriate upper bounds GU and W U for the constraint approach can be speciﬁed. For the region where W (X) has a small value, W (X) is minimized as follows under

Optimization of Spatial Trusses and Frames

277

constraint on G(X): Minimize

W (X)

subject to G(X) ≤ G

(6.16a) U

(6.16b)

X ≤X≤X L

U

(6.16c)

In contrast, for the region where G(X) has small values, G(X) is minimized as follows under constraint on W (X): Minimize

G(X)

subject to W (X) ≤ W

(6.17a) U

X ≤X≤X L

6.4.4

(6.17b) U

(6.17c)

Sensitivity analysis with respect to control points

Problems (6.16) and (6.17) can be solved using a gradient-based optimization algorithm. Therefore, in this section, the sensitivity coeﬃcients with respect to the design variables Xi and Yi are derived for the objective and constraint functions. The sensitivity coeﬃcients of the static displacements are not presented here, because they can be found using the standard methods in Sec. 2.2. If node k is a lower node, the sensitivity coeﬃcients of its coordinates can be obtained directly from (6.7) as C C ∂x(uk ) ∂y(uk ) = Bin (uk ), = Bin (uk ) ∂Xi ∂Yi

(6.18)

In the following, the argument uk is omitted for simplicity. For the upper nodes, the sensitivity coeﬃcients of the oﬀset vector should be computed. Because the vertical oﬀset vector is constant, we derive the sensitivity coeﬃcients of only the normal oﬀset vector. By diﬀerentiating (6.8) and (6.9), and using (6.7), we obtain C

∂n ∂B n =T i ∂Xi ∂u ( ) ∂n0 1 ∂n 1 ∂n =√ + ⊤ n0⊤ n ∂Xi n n ∂Xi n⊤ n ∂Xi

(6.19a) (6.19b)

For the shape deviation, the following equations are derived from (6.14): ] ∫ 1 n [ ∑ C C ∂G ej ) =2 (Xj − X Bin (u)Bjn (u)du , ∂Xi 0 j=1 C

] ∫ 1 n [ ∑ ∂G nC nC e =2 (Yj − Yj ) Bi (u)Bj (u)du ∂Yi 0 j=1 C

(6.20)

278

Optimization of Finite Dimensional Structures 35

Minimum compliance

Shape deviation

30 25 20 15 10 5

Minimum shape deviation

0 0.0002 0.0004 0.0006 0.0008 0.001 0.0012 0.0014 Compliance

FIGURE 6.14: y-coordinates.

Pareto optimal solutions for vertical oﬀset vector, variable

Note that the integration of basis functions has already been carried out in the process of evaluating G(X).

6.4.5

Numerical examples

Consider a 71-bar plane arch-type frame shown in Fig. 6.13 deﬁned using the B´ezier curve of order 6; i.e., the number of control points for deﬁning the lower curve is 7. Because the control points at both ends are ﬁxed at the supports, the number of variable points is 5. The members are rigidly connected at the joints and pin-jointed at the two supports. The span length is 30 m, and the length d of the oﬀset vector is 2 m. The conﬁguration in Fig. 6.13 is assumed to represent the shape preferred by the designer. Pareto optimal solutions are found for two cases where the y-directional coordinates and (x, y)-coordinates of the control points, respectively, are considered as design variables. Therefore, the numbers of design variables are 5 and 10 for these two cases, respectively. The vertical and normal oﬀset vectors are considered for both cases. Optimal shapes are found by using SNOPT Ver. 7.2 (Gill, Murray, and Saunders 2002), where sequential quadratic programming is used. The members consist of a circular tube section with external and internal radii of 50 mm and 46 mm, respectively. The arch is subjected to static loads 2 that represent the self-weight. The elastic modulus is 210 kN/mm , and the 3 weight density is 77.0 kN/m . The bounds for X are not given. In the fol-

Optimization of Spatial Trusses and Frames

279

(a) Minimum G

(b) W = 0.0011

(c) W = 0.0008

(d) W = 0.0005

(e) W = 0.00033

(f) G = 17.5

(g) G = 25.0

(h) G = 32.5

(i) Minimum W

FIGURE 6.15: coordinates.

Optimal shapes for the vertical oﬀset vector, variable y-

(a) Minimum G

(b) W = 0.0010

(c) W = 0.0008

(d) G = 3

(e) G = 6

(f) G = 9

(g) G = 10

(h) G = 11

(i) Minimum W

FIGURE 6.16: coordinates.

Optimal shapes for the normal oﬀset vector, variable (x, y)-

lowing, the units of length and force are m and kN, respectively, which are omitted for brevity. The Pareto optimal solutions are ﬁrst found for the vertical oﬀset vector with only y-coordinates as variables. If the shape deviation G is minimized, the optimal value is 0 and the corresponding value of compliance W is 1.3283× 10−3 . Since the preferred shape is deﬁned using the B´ezier curve of order 6,

280

Optimization of Finite Dimensional Structures

exactly the same shape as shown in Fig. 6.13 is obtained by minimizing G. On the other hand, if W is minimized, the optimal value of W is 3.0099 × 10−4 , and the corresponding value of G is 35.545. We apply the constraint approach for generating the set of Pareto solutions. Problem (6.16) is solved ﬁrst with the upper bound for W decreased as 0.0013, 0.0012, . . . , 0.0004, 0.00035, 0.00033, 0.00032. Then Problem (6.17) is solved with the upper bound for G decreased as 35.0, 32.5, . . . , 15.0. The Pareto optimal solutions are plotted in the objective function space in Fig. 6.14. As is seen, the Pareto solutions form a convex curve in the objective function space. Figs. 6.15(a)–(i) show the shapes of various Pareto solutions. We can see from these ﬁgures that a smaller compliance leads to a circular shape. However, the lengths of the members near the supports become very large if the compliance is minimized with y-directional variables. If (x, y)-coordinates are considered as variables, the Pareto solutions in the objective function space are almost the same as those in Fig. 6.14. Therefore, no signiﬁcant eﬀect is found for considering x-coordinates of the control points as additional variables, if the vertical oﬀset vector is used. Pareto solutions are next generated with the normal oﬀset vector and (x, y)coordinates as variables. If the shape deviation G is minimized, the optimal value is 0 and the corresponding value of compliance W is 1.2541 × 10−3 , which is a little smaller than the case in which the vertical oﬀset vector is considered with y-directional variables. On the other hand, if W is minimized, the optimal value of W is 5.4461 × 10−4 , and the corresponding value of G is 11.956, which shows that the minimum compliance is larger than that for the vertical oﬀset vector even if the (x, y)-coordinates are considered as variables. The shapes of Pareto solutions are shown in Figs. 6.16(a)–(i). As is seen, the upper and lower chords around the center, as well as the members near the supports, become shorter as a result of minimization of compliance with variable (x, y)-coordinates.

6.5 6.5.1

Shape optimization of a single-layer latticed shell Introduction

In Sec. 6.4, we presented an approach to shape optimization of plane archtype frames considering the designer’s preference. The smoothness of the surface, as well as the mesh pattern, is also important for designing latticed shells with non-standard shapes. Complex surfaces deﬁned using parametric surfaces are often called free-form surfaces (e.g., Guillet, No¨el, and L´eon 1996), and the shapes of shell roofs deﬁned using parametric surfaces are often called free-form shells (Kimura and Ohmori 2008). Ramm, Bletzinger, and Reitinger (1993) used B´ezier surfaces for shape opti-

Optimization of Spatial Trusses and Frames R03

R02

R12

281

R04 R13

R11

R22

R21

R01

υ R00

u

R10

R31

R20 R30

R40

FIGURE 6.17: Relation between the control net and nodal locations of the surface deﬁned using a triangular B´ezier patch.

mization of shells. Ohsaki and Hayashi (2000) extended the fairness metrics of surface (Rando and Roulier 1991; Roulier and Rando 1994) and deﬁned some roundness metrics for optimization of ribbed shells. Ohsaki, Nakamura, and Kohiyama (1997) carried out shape optimization of a double-layer roof truss, where the triangular B´ezier patch is utilized for modeling smooth surfaces. In this section, a formulation of the optimization problem and optimization results are presented for single-layer latticed shells modeled using a triangular B´ezier patch. Pareto optimal solutions considering global properties of the surface deﬁned by the variance of member lengths and compliance against static loads, respectively, representing geometrical and mechanical properties, can be successfully generated without loss of smoothness of the surface by modifying the control points of the parametric surface.

6.5.2

Description of a latticed shell and formulation of the optimization problem

The surface of the single-layer latticed shell is deﬁned using the B´ezier triangle, or triangular B´ezier patch, as shown in Fig. 6.17, where the thick lines represent the control net, and the thin lines are members of the frame. The points Rαβ (α = 0, . . . , nC ; β = 0, . . . , nC − α) are the control points, where nC is the order of the surface that is 3 in Fig. 6.17 (see Appendix A.6 for details of the triangular B´ezier patch). Let u and v denote the parameters for the surface, which are combined to a vector u = (u, v)⊤ . The parameter vector of the point at which the kth node is located is denoted by uk = (uk , vk )⊤ ; i.e., the coordinates of the kth node are deﬁned as Pk = P(uk ). The thin lines in Fig. 6.17 are obtained from the

282

Optimization of Finite Dimensional Structures C

30m

20m

x A

B

25m y

(a) Plan

12 m 8m

(b) Elevation (y-direction) FIGURE 6.18:

(c) Elevation (x-direction)

A single-layer latticed dome.

(a) Plan

(b) Elevation (y-direction)

(c) Elevation (x-direction) FIGURE 6.19:

B´ezier patch of the single-layer latticed dome.

Optimization of Spatial Trusses and Frames

283

uniformly spaced parameters 0.0, 0.05, 0.1, . . . , 0.9, 0.95, 1.0 for uk and vk . This way, a single-layer frame with triangular mesh can be generated. An example of a single-layer latticed dome is shown in Fig. 6.18, and the locations of the control points are plotted in Fig. 6.19. The frame is supported at the three corners A, B, and C, at which the control points exist owing to the basic property of the B´ezier surface. Therefore, the locations of the supports can easily be ﬁxed during shape optimization with the coordinates of the control points as design variables. Let K and U denote the stiﬀness matrix and the displacement vector under the speciﬁed load vector, respectively. The vector consisting of the variable coordinates of the control points is denoted by X. Then K and U are functions of X. The compliance W (X) deﬁned as W (X) = U⊤ (X)K(X)U(X) is to be minimized in the following optimization problem so that the frame has suﬃcient stiﬀness against the external loads. Let Li (X) denote the length of the ith member of the frame consisting of m members. The average value of Li (X) among all members is denoted by Lave (X). The variance D(X), deﬁned as follows, of the member lengths is to be minimized so as to improve the regularity of the frame, which leads to eﬃciency in aesthetic and constructional aspects: D(X) =

m ∑

(Lave (X) − Li (X))2

(6.21)

i=1

Hence, we have two objective functions to be minimized, and the optimization problem turns out to be a multiobjective programming (MOP) problem (see Appendix A.4 for basic properties and methodologies of MOP). In order to assure regularity and smoothness of the surface and frame, upper and lower bounds are to be given for the locations of the control points and the nodes of the frame. Let P(X) denote the vector consisting of all components of Pk (X) (k = 1, . . . , np ), where np is the number of nodes of the frame. Then the geometrical constraints are given as His (X) ≤ 0, (i = 1, . . . , ns ) Hif (P(X)) ≤ 0, (i = 1, . . . , nf )

(6.22a) (6.22b)

where ns and nf are the numbers of geometrical constraints for the surface and the frame, respectively, and the simple bound constraints for X and P(X) are assumed to be included. Then the MOP problem is formulated as Minimize W (X) and D(X) subject to His (X) ≤ 0, (i = 1, . . . , ns ) Hif (P(X)) ≤ 0, (i = 1, . . . , nf )

(6.23a) (6.23b) (6.23c)

In the following examples, Pareto optimal solutions are obtained using the constraint approach, and each single-objective optimization problem is solved

284

Optimization of Finite Dimensional Structures

utilizing a gradient-based approach. Therefore, equations are derived below for calculating sensitivity coeﬃcients of the nodal coordinates with respect to the coordinates of the control points. The sensitivity coeﬃcients of the compliance can be found from those of nodal coordinates directly using the formulations in Secs. 2.2 and 2.7. The jth component (j = 1, 2, 3) of Rαβ and Pk in the three-dimensional space is denoted by Rαβj and Pkj , respectively. The following expression is derived from (A.136a) and (A.136b) in Appendix A.6: C ∂Pkj nC ! = uα v β (1 − uk − vk )n −α−β , ∂Rαβj α!β!(nC − α − β)! k k

(6.24)

(j = 1, 2, 3; k = 1, . . . , np ) Note that ∂Pkj /∂Rαβi = 0 for j ̸= i. The sensitivity of the length Lk of the kth member connecting nodes r and s with respect to Rαβj is calculated from ∂Lk ∂Lk ∂Prj ∂Lk ∂Psj = + ∂Rαβj ∂Prj ∂Rαβj ∂Psj ∂Rαβj

6.5.3

(6.25)

Numerical examples

Optimal shapes are found for a single-layer latticed frame, as shown in Fig. 6.18, which has three supports and is modeled using a triangular B´ezier surface of order 6; hence, the number of control points is 28. The height of the supports A, B, and C are 0, 0, and 8 m, respectively. The height of the node indicated by a blank square in Fig. 6.18 is constrained to be 12 m. The locations of the control points at the supports are ﬁxed. Therefore, we have 25 free control points. The parameters u and v are divided into 18 uniformly spaced intervals between 0 and 1 to deﬁne the nodal locations; i.e., uk , vk = 0, 1/18, . . . , 17/18, 1. Optimum designs are found using SNOPT Ver. 7.2 (Gill, Murray, and Saunders 2002). The initial locations of the control points are illustrated in Fig. 6.19. The frame is subjected to static loads that represent the self-weight. The members consist of a circular tube section with external and internal radii of 50 mm 2 and 46 mm, respectively. The elastic modulus is 210 kN/mm , Poisson’s ratio for computing the stiﬀness for uniform torsion is 0.3, and the weight density 3 is 77.0 kN/m . The bounds for X are not given. In the following, the units of length and force are m and kN, respectively, which are omitted for brevity. Consider ﬁrst the case in which only the z-coordinates of the free control points are considered as design variables; i.e., the number of variables is 25. We use the constraint approach in a similar manner as in Sec. 6.4 (see Appendix A.4 for details of the constraint approach). Single-objective problems are ﬁrst solved for minimizing the compliance and the variance of member lengths, respectively, to ﬁnd the admissible upper-bound values for the constraint approach. If the compliance is minimized, the optimal value is

Optimization of Spatial Trusses and Frames

(a) Plan

(b) Elevation (y-direction)

(c) Elevation (x-direction)

(d) Deformation (×1000)

FIGURE 6.20: coordinates.

285

Optimal shape for minimizing compliance; variable z-

W = 0.01522, and the variance of the lengths is D = 59.74. The optimal shape and deformation scaled by a factor of 1000 are plotted in Fig. 6.20, where the dotted lines in Fig. 6.20(d) represent the undeformed shape in Fig. 6.20(b). As is seen, the smooth and round shape is obtained by minimizing W , and the deformation is very small even with the scale factor of 1000. On the other hand, if the variance of the lengths is minimized, the optimal value is D = 23.39, and the compliance is W = 10.51, which is very large. The optimal shape and deformation multiplied by 2 are plotted in Fig. 6.21. As is seen, the smoothness of the surface is deteriorated by minimizing deviation of member lengths. In fact, the principal curvature has diﬀerent signs at the center and a corner. In order to ﬁnd the Pareto optimal solutions, the upper bound of W for the minimization problem of D is reduced from 0.12 to 0.03 with the decrement 0.01. Then, W is minimized under constraint on D, where the upper bound of D is reduced from 59 to 32 with the decrement 3. The Pareto optimal solutions, which form a convex curve in the objective function space, are plotted in Fig. 6.22. The solution for D = 32 is shown in Fig. 6.23 with the deformation scaled by 500. This way, an intermediate shape between Figs. 6.20 and 6.21 that has moderately small deformation can be successfully found using the constraint approach. Next we consider (x, y, z)-coordinates of the free control points as design variables; i.e., the number of design variables is 25 × 3 = 75. In this case, the variance of the member lengths can be minimized to D = 7.513 × 10−8 , which is almost equal to zero, and the compliance of this solution has a very large

286

Optimization of Finite Dimensional Structures

(a) Plan

(b) Elevation (y-direction)

(c) Elevation (x-direction)

(d) Deformation (×2)

FIGURE 6.21: z-coordinates.

Optimal shape for minimizing deviation of lengths; variable 65 60

Variance of length

Minimum compliance 55 50 45 40 35 Minimum variance 30 25 0

0.02

0.04

0.06

0.08

0.1

0.12

0.14

Compliance

FIGURE 6.22:

Pareto optimal solutions; variable z-coordinates.

value, 10.94. The optimal solution that minimizes D and its deformation in real scale are shown in Fig. 6.24. As is seen, a cylindrical surface is generated by minimizing D, which results in a very large deformation. In contrast, the problem of minimizing the compliance did not converge. The constraint approach is also used here for ﬁnding the Pareto optimal solutions. The

Optimization of Spatial Trusses and Frames

(a) Plan

(b) Elevation (y-direction)

(c) Elevation (x-direction)

(d) Deformation (×500)

FIGURE 6.23:

287

Optimal shape for D = 32; variable z-coordinates.

(a) Plan

(b) Elevation (y-direction)

(c) Elevation (x-direction)

(d) Deformation)

FIGURE 6.24: Optimal shape for minimizing deviation of lengths; variable (x, y, z)-coordinates.

solution corresponding to W = 0.02 and its deformation scaled by 500 are shown in Fig. 6.25. This way, a moderately cylindrical shape with small deformation can be found by constraining W to a small value.

288

Optimization of Finite Dimensional Structures

(a) Plan

(b) Elevation (y-direction)

(a) Elevation (x-direction)

(b) Deformation (×500)

FIGURE 6.25:

6.6 6.6.1

Optimal shape for W = 0.02; variable (x, y, z)-coordinates.

Configuration optimization of an arch-type truss with local geometrical constraints Direct assignments of geometrical constraints

So far, we considered shape optimization under constraints on global geometrical measures, e.g., deviation from target shape and variance of member lengths. In this section, a shape optimization method with direct assignment of local geometrical constraints is presented for an arch-type truss modeled using a B´ezier curve, which is optimized utilizing a genetic algorithm (GA) (Ohsaki and Kato 1997). We also optimize the topology in addition to the geometry; i.e., optimal conﬁguration is to be found. Consider an arch-type truss with triangular units (Warren truss), as shown in Fig. 6.26, where the lower nodes exist along the B´ezier curve PL (u) = (PLx (u), PLy (u))⊤ . The parameter u (0 ≤ u ≤ 1) is discretized to nD + 1 points to deﬁne the nodes and supports on the lower curve: uiL =

i , (i = 0, 1, . . . , nD ) nD

(6.26)

where i = 0 and nD correspond to the two supports. Let d denote the distance between the lower and upper curves. Then, the upper curve is given by the oﬀset curve with normal distance d in a similar

Optimization of Spatial Trusses and Frames

FIGURE 6.26:

289

A Warren arch truss. PU (u ) PL (u ) d

FIGURE 6.27: with length d.

PL,u(u )

Deﬁnition of the upper curve using the normal oﬀset vector

c

c

e g

a

e g

a

d

d

b

f (a) Triangles

FIGURE 6.28:

b

f (b) Diagonals

Feasible topology of a Warren truss.

manner as in Sec. 6.4: ( x ) ( )( x ) d PL,u PL (u) 0 −1 +√ PU (u) = y y PL,u PLy (u) 1 0 x )2 + (P 2 (PL,u L,u )

(6.27)

where ( · ),u denotes diﬀerentiation with respect to u; see Appendix A.6.3 for details of the oﬀset curve. Fig. 6.27 illustrates the relation between the upper and lower curves. The parameter values uiU for deﬁning the upper nodes of a Warren truss with additional nodes above the supports are deﬁned as 1 i−1 (u + uiL ), (i = 1, . . . , nD ) 2 L P nP +1 u0U = u0L , uU = unL +1 uiU =

(6.28)

Then, the location of an upper node is given as shown in Fig. 6.27 by incorporating u = uiU into PU (u) in (6.27). We next deﬁne the geometrical parameters rj and ak for constraining the local properties. For example, for the truss in Fig. 6.28(a), there exist ﬁve triangles (a,b,c), (b,c,d), (c,d,e), (d,e,f), and (e,f,g). The sequence of diagonals is shown in Fig. 6.28(b). Let Pa , . . . , Pg denote the location vectors of points ‘a’,. . . ,‘g’. Suppose that these vectors are in the three-dimensional space with

290

Optimization of Finite Dimensional Structures β d

c a

c g

b

a

f e (a) Triangles

d

g f

b

e

(b) Diagonals

FIGURE 6.29: Infeasible topology that does not correspond to a Warren truss.

sk lk

FIGURE 6.30: Deﬁnition of the arc of the lower curve formed by a lower chord. the third component equal to 0. For the third triangle (c,d,e), for example, deﬁne a vector 1 R3 = (Pd − Pc ) × (Pe − Pd ) (6.29) 2 The vectors R1 , R2 , R4 , and R5 are deﬁned similarly using the vector product between the direction vectors of the diagonals. Then, deﬁne rj as the third component of Rj , which is equal to the signed area of the triangle formed by a pair of diagonals and a lower or upper chord. The regularity of the Warren truss is assured through assignment of the constraint on rj . To preserve the conﬁguration of the Warren truss, rj < 0 should be satisﬁed if the triangle contains a lower chord, and rj > 0 if the triangle contains an upper chord. The truss in Fig. 6.28 satisﬁes these conditions. However, for the truss in Fig. 6.29, r3 < 0 for the triangle (c,d,e) containing the upper chord (c,e). Accordingly, the angle β in Fig. 6.29(b) exceeds π, which leads to an infeasible conﬁguration as a Warren truss. Note that the regularity of the truss is maintained by assigning stricter nonzero lower and upper bounds for rj . Furthermore, the distance between the lower curve and the lower chord can be restricted using the length lk of the kth lower chord and the arc length sk of the B´ezier curve associated with the kth lower chord, as shown in Fig. 6.30. The parameter ak is then deﬁned as ak =

sk − lk sk

A lower chord closely follows the curve if ak is close to 1.

(6.30)

Optimization of Spatial Trusses and Frames R1

291

R2

0000 0001 0010 0011 0100 0101 0110 0111

2000 mm

1000 1001 1010 1011

2000 mm

R0 2000 mm

2000 mm

1100 1101 1110 1111

R3

300 mm 300 mm

7500 mm (a)

FIGURE 6.31:

6.6.2

(b)

Assignment of control points.

Optimization using a genetic algorithm

Optimal conﬁguration is found using a genetic algorithm. The objective function to be minimized is the compliance W ; see Secs. 1.8 and 6.4 for the deﬁnition of compliance. We use the B´ezier curve of order 3, for simplicity, considering symmetry conditions; see Appendix A.6 for details. Note that the locations of the supports are ﬁxed. Therefore, we have only one control point with a variable location. For a Warren truss as in Fig. 6.26, we have nD lower chords, nD − 1 upper chords, and 2nD diagonals. Therefore, there are 2nD − 1 triangle units, and the constraints are given as rL ≤ rk ≤ rU , (k = 1, . . . , 2nD − 1)

(6.31a)

ak ≤ a , (i = 1, . . . , n )

(6.31b)

U

D

where the superscripts ( · )L and ( · )U indicate lower and upper bounds, respectively. Note again that rL and rU have positive values if the triangle contains an upper chord, and they are negative if the triangle contains a lower chord. A simple genetic algorithm is used for optimization. Fig. 6.31 illustrates the coding scheme for modeling 16 candidate locations of the control points R1 and R2 , respectively, using a 4-bit string. A binary string of nD −1 bits is used for representing by 1 and 0 the existence and nonexistence, respectively, of the lower node at each candidate point on the lower curve. Fig. 6.32 illustrates the various topologies of a truss with two lower nodes deﬁned from four candidate nodes excluding the supports; i.e., N D = 5. An exterior penalty function approach, presented in Appendix A.2.2.4, is used for incorporating the constraints into the objective function. The penalty

292

Optimization of Finite Dimensional Structures

(a) 0101 (c) 0011 (b) 1001 FIGURE 6.32: Various topologies of a truss with two lower nodes selected from four candidate nodes represented by 4-bit string. function dr (rk ) for the constraint (6.31a) d (r ) = 0 r k )2 ( L r − rk dr (rk ) = τ rU − rL ( )2 rk − r U dr (rk ) = τ rU − rL

is deﬁned as for rL ≤ rk ≤ rU for rk < rL

(6.32)

for rU < rk

where τ is the penalty parameter that should be appropriately assigned in view of the magnitudes of the objective function W and the penalty terms. Let np denote the size of the population (number of individuals in a generation). The binary string of the jth individual of the ith generation is denoted by Xij (j = 1, . . . , np ), and let Dr0 (Xij ) be deﬁned as Dr0 (Xij ) =

D 2n −1 ∑

dr (rk )

(6.33)

k=1

Then, the total penalty Dr (Xij ) for constraint (6.31a) is given as { Dr (Xij ) = Dr0 (Xij ) for Dr0 (Xij ) ≤ D∗ j j ∗ 0 ∗ Dr (Xi ) = D + α(Dr (Xi ) − D ) for Dr0 (Xij ) > D∗

(6.34)

where D∗ and α (< 1) are the parameters, which are given so that the individual with a large violation of constraints does not have too large a penalty term. Various deﬁnitions of penalty functions have been proposed for GAs. For instance, Chen and Chen (1997) deﬁned the penalty parameter τ (k g ), which depends on the generation number k g , as τ (k g ) = τ1 [1 + 0.2(k g − 1)]

(6.35)

where τ1 is the speciﬁed initial value. This way, the penalty parameter linearly increases as the generation proceeds. A linear penalty with a threshold value can also be used, as suggested by Hajela and Lee (1995). The penalty function Da (Xij ) for constraint (6.31b) is deﬁned similarly, and the performance measure C(Xij ) for Xij is obtained by adding the penalty terms to the compliance W (Xij ) as C(Xij ) = W (Xij ) + Dr (Xij ) + Da (Xij )

(6.36)

Optimization of Spatial Trusses and Frames

293

... 1 0 0 1 0 ...

... 1 0 0 0 1 ...

FIGURE 6.33:

... 1 0 1 0 0 ...

Illustration of the mutation process.

FIGURE 6.34: Optimal conﬁguration and stresses with geometrical constraints.

Since GA maximizes the ﬁtness function, the objective function C(Xij ) to be minimized can be transformed to C0 − C(Xij ), with C0 being a suﬃciently large value, or to 1/C(Xij ). However, in order to avoid ambiguity in the transformation of the ﬁtness function, a ranking strategy with linear scaling is used here; i.e., the individual with rank N has the following ﬁtness value f (N ): f (N ) = f w + (f b − f w )

np − R np − 1

(6.37)

where f w and f b are the parameters for deﬁning the maximum and minimum ﬁtness values, respectively. The two-point crossover is used with probability P c . Mutation is deﬁned such that a lower node moves to the right or left with the probability P m . If the number of candidate lower nodes is far larger than that of the existing nodes in the optimal solutions, it is unlikely that there exists a node at one of the two neighboring candidate nodes. Therefore, the number of lower nodes is likely to be preserved and local search can be carried out by mutation, as illustrated in Fig. 6.33. An elitist strategy is used, where only the best individual remains in the next generation without carrying out mutation or crossover.

294

Optimization of Finite Dimensional Structures

FIGURE 6.35: Optimal conﬁguration and stresses without geometrical constraints.

6.6.3

Numerical examples

Optimization is carried out for a symmetric truss subjected to static loads. The depth d is 0.5 m and the span length is 7.5 m. The 16 candidate curves are generated from the location of the control point R1 coded with a string of 4 bits as plotted in Fig. 6.31, where the location of R2 is determined using the symmetry condition. The symmetrically located 100 candidate nodes are described by the string of 50 bits. Therefore, the total number of bits for each individual (solution) is 4 + 50 = 54. The concentrated loads corresponding to the nonstructural mass are applied at the upper nodes, where the distributed mass is 100.0 kg/m for the curve 2 associated with the upper nodes, and the acceleration of gravity is 9.8 m/s . 3 3 The self-weight is also considered with the mass density 7.86 × 10 kg/m . −4 2 The cross-sectional areas are 5.0 × 10 m for all members, and the elastic 2 modulus is 205.8 kN/mm . The bounds for the constraints are (rL , rU ) = (0.4, 1.4) for a triangle containing an upper chord, (rL , rU ) = (−1.4, −0.4) for a triangle containing a lower chord, and aU = 0.05. The parameters for the penalty function are D∗ = 30.0, τ = 8000.0, and α = 0.1. The probabilities for crossover and mutation are P c = 0.6 and P m = 0.02, respectively, and the size of the population is np = 50. The ﬁtness parameters f b and f w for scaling are given so that the ﬁtness of the best solution is 8/3 of that of the worst solution. The optimal solution with W = 138.08 Nm and nine lower chords is found, as shown in Fig. 6.34, where the width of each member is proportional to the absolute value of the stress. If the geometrical constraints are not considered, the optimal solution has only three lower chords, as shown in Fig. 6.35, and the compliance is 113.00 Nm, which is smaller than that for the optimal solution with geometrical constraints. This way, the regularity of the truss can be preserved by sacriﬁcing stiﬀness against static loads.

Optimization of Spatial Trusses and Frames

6.7 6.7.1

295

Seismic design for spatially varying ground motions Introduction

Due to the increasing demand for constructing long-span dome structures, the spatial variation of seismic motions has become an important factor in the ﬁeld of civil and architectural engineering. The simplest among the effects of spatial variation is the wave passage eﬀect, as illustrated in Fig. 6.36, due to the delay of the wave being transmitted to the supports. This eﬀect has been extensively investigated since the 1970s, mainly for the design and construction of bridges. There are other causes of spatial variation of seismic motions, i.e., the incoherency eﬀect due to reﬂections and refractions of the wave in the heterogeneous medium, and the local eﬀect or site response eﬀect due to diﬀerences in soil conditions near the supports. The response of a structure to spatially varying seismic motions may be computed by using time-history analysis against diﬀerent forced displacements at each support (Clough and Penzien 1975; Price and Eberhard 1998), or a frequency domain analysis (DebChaudhury and Gazis 1988; Harichandran, Hawwari, and Swedian 1996). It will be convenient, however, in design practice, if a response spectrum approach can be used, because the characteristics of the ground motions are usually given in the form of a design response spectrum. Der Kiureghian and Neuenhofer (1992) presented a response spectrum approach incorporating the wave passage eﬀect and incoherency eﬀect that is modeled using coherency functions between the ground motions at the supports. The correlations between the modal responses are also considered in a similar manner as the complete quadratic combination (CQC) method (Wilson, Der Kiureghian, and Bayo 1982). Hao and Duan (1995) showed that incoherency of the support motions leads to torsional responses of frames. Zembaty (1996) carried out a parametric study to determine the eﬀect of incoherency parameters on structural responses. In this section, the method of Ohsaki (2001b) is summarized to carry out optimization and postoptimal analysis of long-span structures considering the incoherency eﬀect, the wave passage eﬀect, and the local ampliﬁcation eﬀect of the ground motions at the supports. It is shown that the second-order sensitivity coeﬃcients of the optimal solution with respect to the parameters characterizing the spatial variation of the seismic motions can be easily obtained if the ﬁrst-order parametric sensitivity coeﬃcients vanish.

6.7.2

Response to spatially varying ground motions

The total degrees of freedom (DOFs) of the internal nodes and supports of a structure are classiﬁed into unconstrained DOFs (UDOFs) and support DOFs (SDOFs). Let x(t) and u(t) denote the vectors of absolute displacements

296

Optimization of Finite Dimensional Structures

Support

Input direction

FIGURE 6.36:

d kl

υapp = v s cos υs

vs

Illustration of the wave passage eﬀect of ground motion.

corresponding to UDOF and SDOF, respectively, which are functions of time t. The stiﬀness matrix, mass matrix, and damping matrix, denoted by K, M, and C, respectively, are divided into the components corresponding to UDOF and SDOF that are indicated by the subscripts ( · )x and ( · )u , respectively. Then, the equations of motion are written as (Clough and Penzien 1975): ) )( ) ( ( )( ˙ ¨ (t) Cx Cxu Mx Mxu x(t) x + ˙ ¨ (t) u(t) u C⊤ M⊤ xu Cu xu Mu ( )( ) ( ) (6.38) Kx Kxu x(t) 0 + = u(t) f (t) K⊤ xu Ku where a dot indicates diﬀerentiation with respect to t, and f is the vector of reaction forces due to the displacements u at the support, which are regarded as forced displacements to the structure. Let xs (t) and xd (t) denote the pseudostatic and dynamic components of x(t); i.e., x = xs (t) + xd (t) (6.39) and xs (t) is deﬁned as

xs (t) = −K−1 x Kxu u(t)

(6.40)

Note that the inverse of Kx is not actually computed in the process of structural analysis; i.e., xs (t) is computed from Kx xs (t) = Kxu u(t). The vector xd is expressed as a sum of the modal components: xd (t) =

n ∑

yi (t)ϕi = Φy(t)

(6.41)

i=1

where n is the number of UDOFs, and ϕi is the ith eigenmode of the structure with ﬁxed SDOF.

Optimization of Spatial Trusses and Frames

297

Let ski (t) denote the modal response to the input uk (t) at the kth SDOF. A vector rk is deﬁned as rk = −K−1 (6.42) x Kxu ik where ik is a vector whose kth component is 1 and the remaining components are 0. Suppose the representative response z(t), e.g., strain of a member, is deﬁned as ⊤ z(t) = q⊤ (6.43) 1 x(t) + q2 u(t) which is rewritten as z(t) =

s ∑ n ∑

bki ski (t) +

k=1 i=1 ⊤ ak = q1 rk + q⊤ 2 ik , ⊤ bki = q1 ϕi βki , (k

s ∑

ak uk (t)

(6.44a)

k=1

(k = 1, . . . , s) = 1, . . . , s; i = 1, . . . , n)

(6.44b) (6.44c)

where s is the number of SDOFs, and βki is deﬁned as follows as the participation factor of the ith mode to the input at the kth SDOF: βki = ϕ⊤ i (Mx rk + Mxu ik )

(6.45)

Note that the eigenmodes are normalized with respect to Mx as ϕ⊤ i Mx ϕi = 1. Assuming that uk is a zero-mean jointly stationary process, the mean square γz2 of z(t) is given as γz2 =

s ∑ s ∑

ak al ρuk ul γuk γul + 2

k=1 l=1 s ∑ s ∑ n ∑ n ∑

+

s ∑ s ∑ n ∑

ak blj ρuk slj γuk γslj

k=1 l=1 j=1

(6.46)

bki blj ρski slj γski γslj

k=1 l=1 i=1 j=1

where γuk , γski , ρuk ul , ρuk slj , and ρski slj are coeﬃcients. For example, the coeﬃcient γski is given as (∫ γski =

∞

−∞

|Hi (iω)| Gu¨k u¨k (ω)dω 2

) 12 (6.47)

√ where ω is the circular frequency, i = −1, Gu¨k u¨k is the power spectrum of the input acceleration u ¨k , and Hi (iω) is the transfer function of the ith mode that is a function of the circular frequency ω; see Der Kiureghian and Neuenhofer (1992) for details of other coeﬃcients. Note that the ﬁrst, third, and second terms in (6.46) represent the pseudostatic response, dynamic response, and their combination, respectively. Consider a response to the horizontal component of a ground motion. The coherency between the ground accelerations u ¨k and u ¨l at kth and lth SDOFs

298

Optimization of Finite Dimensional Structures

is given as (Luco and Wong 1986; Zerva 1990; Abrahamson, Schneider, and Stepp 1991). [ ( )2 ] ( ) ηωdLkl αωdkl γkl (iω) = exp − exp i (6.48) vs vapp where α is the incoherency factor, dkl is the horizontal distance between the kth and lth SDOFs, as illustrated in Fig. 6.36, dLkl is the projected distance of dkl to the horizontal plane, vs is the velocity of shear wave, and vapp is the apparent velocity of the shear wave that is given by the angle ψ of the wave direction to the horizontal plane as vapp = vs / cos ψ. In (6.48), the ﬁrst and second exponential terms represent the incoherency eﬀect and the wave passage eﬀect, respectively. Note that η is an auxiliary parameter for indicating incorporation and non-incorporation of the wave passage eﬀect with η = 1 and 0, respectively. If γkl (iω) and Gu¨k u¨k (ω) are given, then Guk uk (ω), Guk ul (iω), Gu¨k u¨l (iω), and Guk u¨k (iω) are easily computed by using the standard formulas of frequency domain analysis, and the integrations are carried out to ﬁnd the mean square response of z(t) deﬁned by (6.46). The ith natural circular frequency and damping ratio are denoted by ωi and hi , respectively. Let z max denote the maximum absolute value of z(t). The ratio of the maximum absolute value to the standard deviation of a zero-mean random process is called the peak factor. Assuming that the peak factors for the input quantities and the responses are the same (Der Kiureghian 1980), the mean maximum response E(z max ) of z max is rewritten from (6.46) as [ s s ∑∑ max umax ak al ρuk ul umax E(z )≃ l k k=1 l=1 s ∑ s ∑ n ∑

+2

ak blj ρuk slj umax SDl (ωj , hj ) k

k=1 l=1 j=1

+

s ∑ s ∑ n ∑ n ∑

(6.49)

1 2

bki blj ρski slj SDk (ωi , hi )SDl (ωj , hj )

k=1 l=1 i=1 j=1

where umax and SDk (ω, h) are the speciﬁed maximum displacement and the k displacement response spectrum of the kth SDOF. The value of SDk (ω, h) varies among the supports due to local soil conditions that characterize the site responses (Der Kiureghian 1996). The power spectrum Gu¨k u¨k (ω) of the input acceleration is deﬁned in terms of the response spectrum. In the numerical examples, the following deﬁnition is used (Der Kiureghian and Neuenhofer 1991, 1992): ( )( )2 4 SDk (ω, h) 2hω ω ck +2 + (6.50) Gu¨k u¨k (ω) = c ck ω k + ω(f)k π πτ ν(ω)

Optimization of Spatial Trusses and Frames

299

where τ is the duration of the motion, ν(ω) is the peak factor of the response to white noise, which, together with ω(f)k , are calculated to represent a ﬁnite power for the pseudostatic input (Der Kiureghian and Neuenhofer 1991; Ohsaki, Tagawa, and Kato 2000; Ohsaki 2001b).

6.7.3

Problem formulation and design sensitivity analysis

Consider a spatial frame with m members. Let Li and Ai denote the length and cross-sectional area of the ith member, respectively. The nodal locations and topology of the structure are ﬁxed. All the cross-sectional properties of each member are assumed to be deﬁned by the cross-sectional area; i.e., the design variables are the cross-sectional areas A = (A1 , . . . , Am )⊤ . The objective function is the total structural volume, and constraints are given for the representative stresses, e.g., the maximum stress at the two edges of the two ends of each member, as illustrated in Fig. 5.2 in Sec. 5.1; i.e., there are four points at which the stress is constrained in each member. v Let σi,j (A) denote the representative mean maximum stress at the jth point of the ith member computed using (6.49), where ak and blk are to be deﬁned appropriately in view of the geometry and topology of the frame. w The stress due to static loads, including self-weight, is denoted by σi,j (A). The optimization problem for minimizing the total structural volume V is formulated as Minimize

V (A) =

subject to σi,j (A)

m ∑

Ai Li i=1 v = σi,j (A)

(6.51a) w U + |σi,j (A)| ≤ σi,j ,

(i = 1, . . . , m; j = 1, . . . , 4) ALi

≤ Ai

(6.51b) (6.51c)

U where ALi is the lower bound for Ai , and σi,j is the upper bound for σi,j (A). Problem (6.51) is a nonlinear programming problem, which is solved in the following examples using a gradient-based approach. Therefore, the design sensitivity coeﬃcients of the objective and constraint functions with respect to Ai are needed. Because the sensitivity coeﬃcients of static stresses are obtained using a well-established method in Sec. 2.2, only the formulations v for σi,j (A) are presented below. It is straightforward to diﬀerentiate (6.40)–(6.42) with respect to Ai ; e.g., (6.42) is ﬁrst rewritten as Kx rk = −Kxu ik (6.52)

which is diﬀerentiated with respect to Ai as Kx

∂Kx ∂Kxu ∂rk =− rk − ik ∂Ai ∂Ai ∂Ai

(6.53)

300

Optimization of Finite Dimensional Structures

Therefore, the right-hand side of (6.53) can be conceived as a static load vector, and we can assume Kx has been already factorized when solving (6.52) for rk . Hence, the inverse matrix of Kx or its sensitivity coeﬃcient is not needed to compute ∂rk /∂Ai . Then, using (6.44b) and (6.44c), the sensitivity coeﬃcients of ak and bki are obtained as ∂q⊤ ∂rk ∂q⊤ ∂ak 1 2 = rk + q⊤ + ik 1 ∂Ai ∂Ai ∂Ai ∂Ai ∂bki ∂q⊤ ∂ϕi ∂βki 1 = ϕ βki + q⊤ βki + q⊤ 1 1 ϕi ∂Ai ∂Ai i ∂Ai ∂Ai

(6.54a) (6.54b)

Note that the sensitivity coeﬃcients of the eigenvalues and eigenmodes are computed utilizing the standard formulas shown in Sec. 2.3. For the participation factor, ∂βki ∂ϕ⊤ i = (Mx rk + Mxu ik ) ∂Ai ∂Ai ) ( ∂rk ∂Mx ⊤ ∂Mxu + ϕ⊤ r + M + ϕ i k x k i i ∂Ai ∂Ai ∂Ai

(6.55)

is derived from (6.45), and the sensitivity coeﬃcients of SDk are found as described in Appendix A.7. The sensitivity coeﬃcients of the transfer function are derived using Ωj = ωj2 as ( ) ∂Hj (±iω) ∂ 1 = ∂Ai ∂Ai ωj2 − ω 2 ± 2ihj ωj ω ( ) ∂Ωj ∂ωj 2 (6.56) = −(Hj (±iω)) ± 2ihj ω ∂Ai ∂Ai ( ) hj ω 2 ∂Ωj = −(Hj (±iω)) 1±i ∂Ai ωj Note that Gu¨k u¨k (ω), Guk u¨l (iω), Guk ul (iω), and γuk do not depend on Ai . From (6.47), we obtain ( ) 21 ∫ ∞ ∂γslj ∂ 1 = ¨l u ¨l (ω)dω 2 2 2 2 Gu 2 2 ∂Ai ∂Ai −∞ (ωj − ω ) + 4hj ωj ω [ ] ∫ ∞ 1 ∂ 1 ∂Ωj = Gu¨ u¨ (ω)dω (6.57) 2 2 2 2 2γslj −∞ ∂Ωj (Ωj − ω ) + 4hj Ωj ω ∂Ai l l { } ∫ 2(Ωj − ω 2 ) + 4h2j ω 2 1 ∂Ωj ∞ =− Gu¨l u¨l (ω)dω 2γslj ∂Ai −∞ [(Ωj − ω 2 )2 + 4h2j Ωj ω 2 ]2 Other terms are diﬀerentiated similarly. Because the terms of partial diﬀerentiation are not included in the integrand, as is seen from (6.57), integration

Optimization of Spatial Trusses and Frames

301

should be carried out only once, and the increase in the number of design variables does not lead to a rapid increase in computational cost.

6.7.4

Postoptimal analysis

Since the solution of the optimization problem (6.51) depends on the parameters that deﬁne the spatial variation of the seismic motion, it is practically important to investigate the sensitivity of the optimal solution and/or the optimal objective value with respect to such parameters. In the following examples, α, η, and the maximum ground displacements are taken as the parameters, and the sensitivity coeﬃcients of optimal solutions with respect to the parameters are computed using the parametric programming approach described in Appendix A.5. Consider, in general, an optimization problem Find

b C(p) = min C(A, p) A

subject to Hi (A, p) ≤ 0, (i = 1, . . . , nI )

(6.58a) (6.58b)

where A and p = (p1 , . . . , pnp )⊤ are the vectors of variables and parameters, respectively, nI is the number of inequality constraints, including the bound b constraints for A, and C(p) is the optimal objective value, which is conceived as a function of the parameter vector p. In the following, a function of p is indicated with a hat as (c · ). Suppose an optimal solution has been found using an appropriate method of nonlinear programming, where the Lagrange multiplier λi for the constraint Hi (A, p) ≤ 0 has also been found as the result of optimization. The sensitivity of the optimal objective value with respect to pj is found in a similar manner as (A.122) in Appendix A.5. The second-order sensitivity coeﬃcients of the optimal objective value with respect to the parameters are given in (A.124). bi /∂pk of the variables and bi /∂pk and ∂ λ Note that the sensitivity coeﬃcients ∂ A multipliers are generally needed for computing the second-order sensitivity b coeﬃcients ∂ 2 C/∂p j ∂pk of the optimal objective value. However, it is not practically acceptable to compute those values for a complex optimization problem as considered in this section, because the Hessian of each constraint bi /∂pk and function with respect to the variables is needed for computing ∂ A b ∂ λi /∂pk . Consider a case where all the variables of the optimal solution are even functions of a parameter pk . In this case, the ﬁrst-order parametric sensitivbi /∂pk vanish at pk = 0 and the second-order bi /∂pk and ∂ λ ity coeﬃcients ∂ A sensitivity coeﬃcients of the optimal objective value with respect to the parameters are obtained from ∑ b ∂2C ∂ 2 Hi ∂2C = + λi ∂pj ∂pk ∂pj ∂pk i=1 ∂pj ∂pk nI

(6.59)

302

Optimization of Finite Dimensional Structures

Therefore, for a symmetric structure subjected to ground motions with symmetry or antisymmetry properties, the ﬁrst-order sensitivity coeﬃcients vanish and the second-order coeﬃcients give useful information on the characteristics of the optimal solutions. The entire process of analysis, sensitivity analysis, optimization, and postoptimal analysis is summarized as follows: Step 1: Deﬁne the geometry and material properties of the structure, and set ak and bki in (6.44) for each response quantity to be constrained. Step 2: Assign the parameters ck , SDk , α, η, vs , dkl , and dLkl , and compute ν and ω(f)k ; see Ohsaki, Tagawa, and Kato (2000) and Ohsaki (2001b) for details. Step 3: Initialize the design variable vector A and optimize the structure as follows: 3.1 Analysis: (a) Carry out eigenvalue analysis of free vibration of the structure with ﬁxed supports. (b) Compute γkl (iω) and Gu¨k u¨k (ω) from (6.48) and (6.50), respectively. (c) Evaluate Gu¨k u¨l (iω), Guk u¨k (iω), etc. from Gu¨k u¨k (ω) and Hi (iω) using the standard formulas of frequency domain analysis. (d) Calculate E(z max ) for each response quantity from (6.49). 3.2 Compute the design sensitivity coeﬃcients of E(z max ) as described in Sec. 6.7.3. 3.3 Modify A in accordance with the optimization algorithm and go to Step 3.1 if not converged. Step 4: Carry out postoptimal analysis.

6.7.5

Numerical examples

Consider a rigidly jointed 39-bar arch-type plane frame, as shown in Fig. 6.37. The span length is 100.0 m and the lower nodes are located along a circle with an open angle of 50 deg. The upper nodes are also on a circle, and the diﬀerence between the radii of the two circles with the same center is 7.5 m. Note that the upper chords, lower chords, and diagonals have the same lengths, respectively. The frame is pin-supported at the two ends. The frame is subjected to a set of horizontal seismic motions compatible with the response spectrum of Newmark and Hall (1982); see Appendix A.7 for the deﬁnition and parameters of the spectrum. In the deﬁnition in Appendix A.7, however, the maximum response displacement at Ωr = 0 does not agree with the maximum displacement CD of the ground motion. Therefore,

Optimization of Spatial Trusses and Frames

303

mass 100 m

FIGURE 6.37:

A 39-bar arch-type plane frame.

FIGURE 6.38: Optimal cross-sectional areas without spatial variation of seismic motion. (5)

(5)

SD (Ωr , hr ) is replaced by SD (Ωr ) = CD (1.0 + κΩr ), and κ is deﬁned to sat(4) (5) isfy SD (Ωr ) = SD (Ωr ) at the prescribed eigenvalue. The value of umax for k the kth SDOF is deﬁned by using the scaling parameter µk to incorporate the local ampliﬁcation eﬀect as umax = µk CD . Note that SDk is also multiplied k by µk . Each member of the frame is assumed to be made of sandwich sections with Ii = r2 Ai , where Ii is the second moment of inertia, and r = 0.5 m is the distance between the ﬂange and the member axis that is considered as a 2 constant. The elastic modulus is 205.8 kN/mm . The upper-bound stress is 2 U given as σi,j = 205.8 N/mm for all four points of members, and the lowerbound cross-sectional area is 10.0 mm2 , which is not active at the optimal solution in the following examples. The mass density of the members is 7.86× 3 103 kg/m . A nonstructural mass of 2.0×104 kg is located at each lower node, and its rotational inertia is neglected. The loads in the vertical direction representing the weights of the members and the nonstructural masses are w applied at the nodes to calculate the static stresses σi,j . The parameters of soil condition and the direction of seismic wave are vs = 400.0 m/s and vapp = 2000.0 m/s. The modal damping ratio of the structure is 0.02, which is assumed to be independent of the frequency, for simplicity. The duration τ of the motion is 25.0 sec. In this case, the value of ω(f)k to satisfy ﬁnite power against the pseudostatic input corresponding to the given response spectrum is 0.258. The parameter ck in (6.50) is 3.0. The library DOT Ver. 5.0 (VR&D 1999) is used for optimization. The 16-point Gaussian quadrature is used for integration for the coeﬃcients γski , etc., in the frequency domain. Fig. 6.38 shows the optimal cross-sectional areas for the case without the

304

Optimization of Finite Dimensional Structures

Objective function V

1.67

1.66

1.65

1.64

1.63 0

0.05

0.1

0.15

0.2

0.25

Incoherency parameter α

FIGURE 6.39: Optimal objective values for α = 0.0, 0.01, 0.1, 0.15, 0.2 (triangular mark), and second-order approximation at α = 0.0 (solid line).

eﬀects of spatial variation of seismic motions; i.e., α = η = 0, where the width of each member is proportional to the cross-sectional area. As is seen, the lower chords near the supports and the upper chords around the center have large cross-sectional areas. The optimal objective value Vb is equal to 1.63454 m3 . The triangles in Fig. 6.39 are the values of Vb for α = 0.0, 0.05, 0.1, 0.15, and 0.2. The solid line shows the second-order approximation at α = 0.0, where ∂ 2 Vb /∂α2 = 1.57076 m3 . Note that Vb is an even function of α; i.e., ∂ Vb /∂α = 0 at α = 0, which is obvious from the absence of a linear term and the existence of quadratic term of α in (6.48). It may be observed from Fig. 6.39 that the optimal objective values are successfully approximated as a quadratic function of α. The optimum design for α = 1.0, shown in Fig. 6.40, has larger crosssectional areas in the upper chords around the center than those for α = 0, because the diﬀerence in the movements of two supports causes pseudostatic deformation, which leads to bending deformation of the arch around the center. The optimal objective value for α = 1.0 is 1.71497 m3 . The value of ∂ Vb /∂α at α = 0.5 is 0.485797 m3 , while the parametric sensitivity coeﬃcient obtained by the central diﬀerence method with ∆α = 0.01 is 0.48 m3 , which agrees with good accuracy with the analytical result. CPU time for postoptimal analysis is only about 0.2% of that for optimization. The optimal cross-sectional areas considering the wave passage eﬀect only, i.e., α = 0 and η = 1, are almost the same as those in Fig. 6.40, where the optimal objective value is 1.69317 m3 . Note that η is an auxiliary integer parameter for indicating incorporation of the wave passage eﬀect. Therefore, variation of η indirectly corresponds to variation of vapp , as observed from the term η/vapp in (6.48). The parametric sensitivity coeﬃcient with respect to η at η = 1.0 is 5.14221 × 10−2 m3 , and the sensitivity coeﬃcient obtained by

Optimization of Spatial Trusses and Frames

FIGURE 6.40: eﬀect (α = 1.0).

305

Optimal cross-sectional areas considering the incoherency

the central ﬁnite diﬀerence approach with ∆η = 0.1 is 5.14 × 10−2 m3 , which agrees with good accuracy with the analytical result. Finally, optimum designs are found considering the diﬀerence in the ampliﬁcation of the seismic motions by local soil. Let µ1 and µ2 denote the scaling factors of the maximum ground displacements and the response spectra corresponding to the horizontal displacements of the two supports. The distribution of optimal cross-sectional areas for (µ1 , µ2 ) = (1.2, 0.8) is similar to that in Fig. 6.40, and the optimal objective value is 1.77694 m3 . Because Vb is an even function of µ1 − µ2 , and an odd function of µ1 + µ2 , the sensitivity coeﬃcients with respect to µ1 and µ2 are the same and equal to 0.493577 m3 , whereas the coeﬃcient by the central ﬁnite diﬀerence approach with ∆µ1 = ∆µ2 = 0.05 is 0.5107 m3 , which is slightly diﬀerent from the analytical result. However, the local eﬀect can be successfully approximated using parametric sensitivity coeﬃcients.

6.8 6.8.1

Substructure approach to seismic optimization Introduction

Spatial structures, such as long-span dome structures and bridges, are often built on supporting (boundary) structure. Therefore, the ﬂexibility of the lower supporting structure should be taken into account when evaluating the seismic responses of the upper roof structure. A response spectrum approach, e.g., the complete quadratic combination (CQC) method, may be eﬀectively used for reducing the computational cost; see Appendix A.7. However, when only the upper structure is modiﬁed, the computational cost for eigenvalue analysis is still large if the whole structure is to be analyzed and the number of degrees of freedom (DOFs) of the lower structure is much larger than that of the upper structure. If the damping properties are modeled based on Rayleigh damping, then a diﬃculty also arises for the case where the upper and lower structures have diﬀerent damping properties, e.g., a steel roof supported by a reinforced concrete structure.

306

Optimization of Finite Dimensional Structures

Consider the process of carrying out seismic response analysis for upper and lower structures independently. In this case, the response of the lower structure can be conceived as an input to the upper structure, and diﬀerent motions may be applied to the connections between the two substructures. This situation is similar to the seismic response analysis considering spatial variation of seismic motions that results in quasistatic forced deformation of the structure; see Sec. 6.7 for details. The dynamic analysis method utilizing division of the structure into several domains is called the substructure approach (Meirovitchand and Hale 1981). If the substructures have hierarchical relations, the substructure supported by the primary structure (PS) is called the secondary structure (SS), which includes nonstructural equipment, a tower-type structure on a building, etc. The dynamic response of the PS can be conceived as the input to the SS, and an SS should be designed to prevent resonance with the PS. Hence, various approaches have been presented for evaluating the seismic response of the SS. The ﬂoor response spectra can be used if the weight of the SS is suﬃciently small compared with that of the PS, and accordingly, the interaction in inertia between the SS and the PS can be neglected (Villaverde 1997). Gupta (1997) developed a method of frequency domain analysis that rigorously considers the interaction between the PS and the SS that is supported at a single point by the PS. Dey and Gupta (1998) extended this method to SS with multiple supports, assuming one-to-one correspondence between the DOFs of the SS and the PS at the connections. In this section, a general and computationally eﬃcient approach by Ohsaki (2003b) is summarized for the seismic design of the SS while the properties of the PS are ﬁxed. The method is applied to the optimum design of arch-type trusses supported by column-type trusses.

6.8.2

Frequency domain analysis for a secondary structure

Consider an arch-type truss, as shown in Fig. 6.41, supported by columntype trusses, which are conceived as the SS and the PS, respectively. Fig. 6.42 illustrates the connection between the PS and the SS, where the solid lines are the connecting members. Let N p and N s denote the numbers of DOFs of the PS and the SS, respectively. In the following, the superscripts ( · )p and ( · )s are used for the values corresponding to the PS and the SS, respectively, for which the displacement vectors are denoted by up = (up1 , . . . , upN p )⊤ and us = (us1 , . . . , upN s )⊤ . Note that the ﬁrst np components of up are the displacements of the nodes of the connecting members on the PS (node ‘a’ in Fig. 6.42). Similarly, the ﬁrst ns components of us are the displacements of the nodes of the connecting members on the SS (nodes ‘b’ and ‘c’ in Fig. 6.42). Therefore, np = 2 and ns = 4 for the connection in Fig. 6.42. Let di denote the displacement vector of the connecting nodes in the SS that is induced by a unit rigid-body translation of the connecting members in

Optimization of Spatial Trusses and Frames

307

L

Secondary structure

Primary structure

H

y

H

x WW

FIGURE 6.41:

An arch-type truss supported by column-type trusses. Secondary structure

u4s Connecting member

u3s

b

u2s

u2p c a

u1s

u1p Primary structure

FIGURE 6.42:

Connection between the PS and the SS.

the direction of up . For the connection in Fig. 6.42, the unit translation in the direction of displacement up1 leads to the unit translations in displacements us1 and us3 ; i.e., d1 = (1, 0, 1, 0)⊤ . The ns × np matrix of which the ith column is di is denoted by D. The rth eigenmode of free vibration of the PS without the SS is denoted pr ⊤ by Φpr = (Φpr 1 , . . . , ΦN p ) . Similarly, the eigenmode of the SS, including the connecting members with ﬁxed supports at the connecting nodes in the PS, is sr ⊤ denoted by Φsr = (Φsr 1 , . . . , ΦN s ) ; i.e., for the example in Fig. 6.42, node ‘a’

308

Optimization of Finite Dimensional Structures

is to be ﬁxed. Let Hrp (ω) denote the transfer function of the rth eigenmode of PS. In the following, the circular frequency ω in the argument indicates a function in the frequency domain. Let usc denote the vector consisting of the ﬁrst ns components of us . The ﬁrst ns columns of the damping and stiﬀness matrices of the SS are denoted by Cs and Ks , respectively. The ﬁrst ns columns of the mass matrix of the SS after removing the connecting members is denoted by Ms . Then, an N s × ns matrix R is deﬁned as R(ω) = ω 2 Ms − iωCs − Ks

(6.60)

where i is the imaginary unit. The interaction force between the PS and the SS can be evaluated from the deformation of the connecting members and R(ω). The Fourier transformation of the seismic acceleration is denoted by z¨(ω). The participation factor of the rth eigenmode corresponding to the speciﬁed input direction is denoted by βrp . The transfer function of the rth mode of the SS to the seismic motion at the ﬁxed base is denoted by Hrs (ω). The np × np stiﬀness matrix of the connecting members is denoted by Kc . Then the Fourier transformation of the displacements usi (ω) (i = 1, . . . , N s ) of the SS can be expressed by z¨(ω) after solving the following set of N s linear equations: ( N s N s ns np N p np ∑ ∑∑∑ ∑ ∑ sl pr pr usi (ω) − Φsl i Φt Φm Φk k=1

l=1 t=1 j=1 m=1 r=1

)

c usk (ω) Hls (ω)Hrp (ω)Rtj (ω)Djm Kmk

[ Ns Ns ∑∑ sl s =− Φsl i Φt Hl (ω)

(6.61)

l=1 t=1

(

Mtts rt

n ∑ N ∑ n ∑ s

+

p

)]

p

p p Rtj (ω)Djm Φpr m Hr (ω)βr

z¨(ω),

j=1 r=1 m=1

(i = 1, . . . , N s ) where the components of matrices are indicated using subscripts. This way, the transfer function hsi (ω) = usi (ω)/¨ z (ω) of usi can be found for seismic motion z¨(ω). The power spectrum Gusi usi (ω) of usi is obtained by multiplying (hsi (ω))2 by Gz¨z¨(ω) for the seismic acceleration. The mean square response can be found by integrating Gusi usi (ω) in an appropriate frequency domain. The speciﬁed peak factor is multiplied by the mean square response to obtain the maximum response in a similar manner as in Sec. 6.7. In the process of the seismic design of structures, a response spectrum, instead of a power spectrum, is usually given. Therefore, for example, (6.50)

Optimization of Spatial Trusses and Frames

309

in Sec. 6.7 can be used for converting the displacement response spectrum SD (ω, h) to the power spectrum (Der Kiureghian and Neuenhofer 1992).

6.8.3

Optimization problem

The cross-sectional areas A = (A1 , . . . , Am )⊤ of the SS are optimized for a given design of the PS, where m is the number of members in the SS. The objective function to be minimized is the total structural volume V (A) of the SS. The direction of seismic motion, the design response spectrum SD (ω, h), and the modal damping ratios are given. Constraints are assigned for the mean-maximum response strain εi (> 0) of the ith member of SS. Let εU denote the upper bound for εi . The optimization problem is formulated as Minimize

V (A)

(6.62a)

subject to εi (A) ≤ ε , (i = 1, · · · , m) U

Ai ≥

ALi ,

(i = 1, . . . , m)

(6.62b) (6.62c)

where ALi is the lower bound for Ai . Optimal cross-sectional areas are found using the following algorithm: Step 1: Specify the properties of the PS and carry out eigenvalue analysis to ﬁnd the natural circular frequencies ωrp , eigenmodes Φpr , and participation factors βrp of the PS. Deﬁne the damping ratio hpr and compute the modal transfer function Hrp (ω) of the PS. Step 2: Specify the duration of the seismic motion and the peak factors for the seismic motion and white noise, and obtain Gz¨z¨(ω) from SD (ω, h) using (6.50). Step 3: Deﬁne the damping ratios of the SS, and assign the initial values of Ai . Step 4: Compute the natural circular frequencies ωrs , eigenmodes Φsr , and modal transfer function Hrs (ω) of the SS with ﬁxed support. Step 5: Solve (6.61) for the transfer function of the responses of the SS against the speciﬁed seismic motion. Multiply the peak factor by the mean-square response strains to obtain the mean-maximum response strain εi . Step 6: Modify A according to the optimization algorithm, and go to Step 4 if not converged. This way, the eigenvalue analysis of the PS should be carried out only once at the beginning of the optimization process. Hence, this algorithm is very eﬀective if the number of degrees of freedom of the PS is much larger than that of the SS.

310

Optimization of Finite Dimensional Structures

FIGURE 6.43: Optimal cross-sectional areas for α = 1.0 and εU = 0.0005.

6.8.4

Numerical examples

Optimum designs are found for an arch-type truss supported by columntype trusses, as shown in Fig. 6.41, where L = 20 m, W = 1 m, H = 4 m, and the open angle of the lower circle of the arch is 50 deg. The upper and lower circles have the same center, and the diﬀerence in their radii is 1.5 m. The lower chords, upper chords, and diagonals have the same lengths, respectively. 2 The elastic modulus is 210.0 kN/mm for both the PS and the SS. Although the method in this section can be used for the case where the PS and the SS have diﬀerent damping properties, the stiﬀness-proportional damping with a damping ratio of 0.02 for the lowest eigenmode is used for both the PS and the SS for comparison with the CQC method (Wilson, Der Kiureghian, and Bayo 1982); see Appendix A.7. Optimization is carried out using IDESIGN Ver.3.5 (Arora and Tseng 1987). A ﬁnite diﬀerence approach is used for sensitivity analysis. The cross-sectional areas of the PS are deﬁned by a parameter α as 0.05α m2 to investigate the dependence of the optimal solution of the SS on the stiﬀness of the PS. The lower bound for Ai of the SS is 0.01 m2 . Concentrated masses of 1.0 × 104 kg and 1.0 × 105 kg are located at the nodes of the SS and the PS, respectively. The member mass is not included for brevity. Only the horizontal motions are considered, and the displacement response spectrum by Newmark and Hall (1982) is used for deﬁning the set of strong seismic motions (see Appendix A.7 for the deﬁnition of the parameter values for the spectrum). The duration of motion is 25.0 s, and the peak factors for both the seismic motion and white noise are 3.0 (Der Kiureghian 1980; Der Kiureghian and Neuenhofer 1992). The optimal cross-sectional areas for α = 1.0 and εU = 0.0005 are as shown in Fig. 6.43, where the width of each member is proportional to its cross-sectional area. The fundamental eigenmode of the SS is antisymmetric with respect to the vertical center axis; hence, the cross-sectional areas of the members around the center are relatively small, because the deformation corresponding to the ﬁrst mode is relatively small in this region. The relations between εU and the optimal value of V for various values of α are plotted in Fig. 6.44. As is seen, V is a decreasing function of εU . For a larger value of α, the axial forces near the connections between the PS and the SS have larger values, and V increases accordingly. Let T1s and T1p denote the fundamental natural periods of the SS and the PS, respectively,

Optimization of Spatial Trusses and Frames

311

Total structural volume

0.6 α=1.0 α=1.2 α=1.4 α=1.6 α=1.8

0.5

0.4

0.3

0.2 0.0006

0.0008

0.001

Upper-bound strain

FIGURE 6.44: Relation between the upper-bound strain and the optimal total structural volume (m3 ).

Ratio of natural period

2 α=1.0 α=1.2 α=1.4 α=1.6 α=1.8

1.8 1.6 1.4 1.2 1 0.8 0.0006

0.0008

0.001

Upper-bound strain

FIGURE 6.45: Relation between the upper-bound strain and the ratios of the fundamental periods of the optimal solutions.

with ﬁxed supports. The relation between εU and the ratio T1s /T1p of the optimal solutions is plotted in Fig. 6.45. As is seen, the ratio decreases as εU is decreased, because the stiﬀness of the SS becomes larger. Note that there are some optimal solutions with T1s /T1p = 1, which is usually to be prevented. This fact indicates that the properties of the responses cannot be discussed based on the lowest eigenmodes only; i.e., the eﬀects of the higher modes should be incorporated. Next, we consider asymmetric input to the arch-type truss through modiﬁcation of the cross-sectional areas AR of the members in the right column-type truss for the case α = 1.0 and εU = 0.0005. The cross-sectional areas of the arch are linked to preserve the symmetry property. The values of V (m3 ) for AR = 0.06 and 0.04 m2 are 0.54250 and 0.56672, respectively, which are 16%

312

Optimization of Finite Dimensional Structures

FIGURE 6.46: AR = 0.06 m2 .

Optimal cross-sectional areas for α = 1, εU = 0.0005, and L

H H

WWWW

FIGURE 6.47: 2).

W W WW

An arch-type truss supported by column-type trusses (S =

and 21% larger than that of the symmetric case with AR = 0.05 m2 . Therefore, larger cross-sectional areas are required for asymmetric cases to satisfy the strain constraints, because the quasistatic component of the response increases as a result of the diﬀerence of the motions at the two connections. The optimal cross-sectional areas for AR = 0.06 m2 are as shown in Fig. 6.46. As is seen, the cross-sectional areas of the upper and lower chords around the center have larger values than those in Fig. 6.43 due to the bending deformation of the arch corresponding to the quasistatic component. To investigate the relation between the DOFs of the PS and the computational cost, optimal solutions are found for diﬀerent numbers of spans S of the PS. The geometry for S = 2 is as shown in Fig. 6.47. The results are compared with those by the completely quadratic combination (CQC) method. Computation was carried out using a PC with AMD Athron 1.0GHz. The optimization results are shown in Table 6.5, where the number of steps indicates the number of design modiﬁcations in the optimization process. The parameters are α = 1.0 and εU = 0.0005. It is seen from Table 6.5 that the mean CPU time for an optimization process (Steps 4–6) does not depend on

Optimization of Spatial Trusses and Frames

313

TABLE 6.5:

Number of steps, total CPU time (s), and mean CPU time (s) for the optimization process of the PS with various numbers of span S. Substructure CQC method S DOF of PS Steps Total Mean Steps Total Mean CPU CPU CPU CPU 1 24 9 25.4 2.44 12 10.1 0.84 2 40 12 35.9 2.74 9 11.3 1.26 3 56 9 27.3 2.69 11 22.1 2.01 4 72 9 28.0 2.76 17 50.9 2.99 5 86 8 25.7 2.82 19 81.6 4.29

S if the substructure method is used, while it is almost proportional to S if the CQC method is used; hence, the substructure method is very eﬀective if the number of DOFs of the PS is very large compared with that of the SS.

Appendix

A.1 A.1.1

Mathematical preliminaries Positive deﬁnite matrix and convex functions

A symmetric n × n matrix B is said to be positive semideﬁnite if b⊤ Bb ≥ 0

(A.1)

for any nonzero n-vector b. The matrix B is positive deﬁnite if (A.1) is satisﬁed with strict inequality. All the eigenvalues of a positive deﬁnite matrix are positive. A set C of n-vector x is said to be convex if x = αx1 + (1 − α)x2 ∈ C

(A.2)

for any parameter value 0 ≤ α ≤ 1 and any pair of n-vectors x1 ∈ C and x2 ∈ C. A function F (x) deﬁned in a convex set C of n-vector x is said to be convex if F (x) ≤ F (αx1 + (1 − α)x2 ) (A.3) for any parameter value 0 ≤ α ≤ 1 and any pair of n-vectors x1 ∈ C and x2 ∈ C. Furthermore, F (x) is strictly convex if (A.3) is satisﬁed with strict inequality. Let F1 (x), . . . , Fp (x) denote convex functions deﬁned in the convex set C. Then the pointwise maximum F max (x) = max{F1 (x), . . . , Fp (x)}

(A.4)

is also a convex function (Boyd and Vandenberghe 2004), as illustrated in Fig. A.1. A function F (x) deﬁned in a convex set C of n-vector x is said to be quasiconvex if F (x2 ) ≤ F (x1 ) =⇒ F (αx1 + (1 − α)x2 ) ≤ F (x1 )

(A.5)

315

316

Optimization of Finite Dimensional Structures

F1(x)

F max (x) F2(x) F3(x)

x

FIGURE A.1: An example of a pointwise maximum of convex functions. F(x)

β

Sβ

x

FIGURE A.2: An example of a quasiconvex function. holds for any parameter value 0 ≤ α ≤ 1. Also, if the level set Sβ deﬁned as Sβ = {x ∈ C | F (x) ≤ β}

(A.6)

is a convex set, then F (x) is quasiconvex, as illustrated in Fig. A.2. Note that a convex function is quasiconvex. A function F (x) deﬁned in a convex set C of n-vector x is said to be pseudoconvex if (x1 − x2 )⊤ ∇F (x2 ) ≤ 0 =⇒ F (x1 ) ≤ F (x2 ) (A.7) Note that a convex diﬀerentiable function is pseudoconvex.

A.1.2

Rayleigh’s principle

Let B denote an n × n symmetric positive deﬁnite matrix, and C be an n × n symmetric matrix. A generalized eigenvalue problem is formulated as CΦi = Ωi BΦi , (i = 1, . . . , n)

(A.8)

Appendix

317

where Ωi and Φi are the ith eigenvalue and eigenvector, respectively, and Ωi is ordered as Ω1 ≤ Ω2 ≤ · · · ≤ Ωn (A.9) The ortho-normalization condition is given as Φ⊤ i BΦj = δij , (i = 1, . . . , n)

(A.10)

where δij is the Kronecker delta. From (A.8) and (A.10), we obtain Φ⊤ i CΦj = δij Ωi , (i, j = 1, . . . , n)

(A.11)

The Rayleigh quotient R(b) is deﬁned for a nonzero n-vector b as R(b) =

b⊤ Cb b⊤ Bb

(A.12)

Suppose b is written as a linear combination of the eigenvectors as b=

n ∑

ci Φi

(A.13)

i=1

with the coeﬃcients ci satisfying n ∑

c2i ̸= 1

(A.14)

i=1

From (A.10)–(A.14), we obtain n ∑

R(b) =

c2i Ωi

i=1 n ∑

(A.15) c2i

i=1

Hence, from (A.9), (A.14), and (A.15), Rayleigh’s principle states that the following inequalities are satisﬁed for any nonzero vector b: Ω1 ≤ R(b) ≤ Ωn

(A.16)

The equalities R(b) = Ω1 and R(b) = Ωn are satisﬁed for b = Φ1 and b = Φn , respectively, and the maximum and minimum eigenvalues Ωn and Ω1 are obtained, respectively, by maximizing and minimizing R(b) with respect to b. Example A.1 Let B and C be given as B=

( ) ( ) 1 0 5 −2 , C= 0 1 −2 2

(A.17)

Optimization of Finite Dimensional Structures

318

Rayleigh quotient R(θ)

7 6 5 4 3 2 1 0 0

1

2

3

Parameter θ

FIGURE A.3: Variation of Rayleigh quotient R(θ) with respect to parameter θ. Then we have 1 Φ1 = √ 5

( ) ( ) 1 1 2 , Φ2 = √ , Ω1 = 1, Ω2 = 6 2 −1 5

(A.18)

Suppose b = (b1 , b2 )⊤ is parameterized by θ as b1 = cos θ, b2 = sin θ

(A.19)

Then the Rayleigh quotient R(θ) is obtained as R(θ) = 5 cos2 θ − 4 sin θ cos θ + 2 sin2 θ

(A.20)

which is plotted in Fig. A.3. As is seen, R(θ) has maximum value 6 and minimum value 1, which are equal to Ω2 and Ω1 , respectively.

A.1.3

Singular value decomposition

Let D denote an n × m rectangular matrix, and suppose n ≤ m for simplicity. The eigenvalues of the m × m symmetric matrix D⊤ D are denoted by Ωi (i = 1, . . . , m), which are nonnegative because D⊤ D is positive semideﬁnite. The number r of nonzero eigenvalues is equal to the rank r of D. Then the singular value decomposition (SVD) of D is written as (Horn and Johnson 1990; Atkinson 1989) D = SΩR⊤ (A.21) (

where Ω=

) diag(ω1 , . . . , ωr ) O O O

(A.22)

Appendix

319

√ Here, O is a null matrix of appropriate size, and the diagonal terms ωi = Ωi of the n × m rectangular matrix Ω are called singular values of D, which are deﬁned in nonincreasing order as ω1 ≥ . . . , ≥ ωr > 0. The m × m matrix R and n × n matrix S are orthogonal matrices satisfying R⊤ R = RR⊤ = Im , S⊤ S = SS⊤ = In

(A.23)

where Im and In are the m × m and n × n identity matrices, respectively. By premultiplying S⊤ and postmultiplying R on both sides of (A.21) and using (A.23), we obtain S⊤ D = ΩR⊤ , DR = SΩ

(A.24)

Hence, the column vectors of S and R are called left and right singular vectors, respectively. For the case r < m, the column vectors Ri (i = r + 1, . . . , m) of R correspond to the zero singular value as DRi = 0

A.1.4

(A.25)

Directional derivative and subgradient

The directional derivative F ′ (x0 , s) of the function F (x) at x = x0 in the direction of s is deﬁned as F (x0 + ts) − F (x0 ) t→0+ t

F ′ (x0 , s) = lim

(A.26)

Let X denote a nonempty open convex set of x, and suppose F (x) is a convex function of x in X . The subgradient ζ of F (x) at x = x0 is deﬁned as a vector satisfying (Rocakfellar 1970; Ekeland and T´emam 1999) F (x) ≥ F (x0 ) + ζ ⊤ (x − x0 ) for all x ∈ X

(A.27)

Although a subgradient is originally deﬁned for a convex function, it can also be deﬁned for a concave function as (Floudas 1995) F (x) ≤ F (x0 ) + ζ ⊤ (x − x0 ) for all x ∈ X

(A.28)

The set of all subgradients of F (x) is called the subdiﬀerential of F (x), which is denoted by ∂F (x).

A.2 A.2.1

Optimization methods Classiﬁcation of optimization problems

A problem of minimizing or maximizing an objective function under some constraints on the functions of the variables is called a mathematical program-

320

Optimization of Finite Dimensional Structures

ming problem or optimization problem. The solution methods for mathematical programming problems are called mathematical programming (Arora 2004; Haftka, G¨ urdal, and Kamat 1990; Luenberger 2003). Although most of the structural optimization problems presented in this book can be solved by simply using optimization libraries or software packages, it is necessary to have a good knowledge of optimization algorithms and the classiﬁcation of problems to select an appropriate method and to understand the output data and messages from the program if it could not produce a reasonable optimal solution. The optimization algorithms are classiﬁed into mathematical programming and heuristics, including genetic algorithms and simulated annealing. We focus on mathematical programming approaches in this section (see Sec. A.3 for details of heuristic approaches). Let x = (x1 , . . . , xm )⊤ denote the variable vector, where m is the number of variables. In the following, a vector is assumed to be a column vector, and its component is indicated by a subscript. The objective function to be minimized is denoted by F (x). The equality constraints and inequality constraints are given as Gi (x) = 0 (i = 1, . . . , nE ) and Hi (x) ≤ 0 (i = 1, . . . , nI ), which can be simply written as G(x) = 0 and H(x) ≤ 0, respectively. Throughout the book, an inequality a ≤ b for vectors a and b of the same size means that inequality is satisﬁed for each pair of components of the vectors. A mathematical programming problem is generally formulated as Minimize

F (x)

(A.29a) E

subject to Gi (x) = 0, (i = 1, . . . , n )

(A.29b)

Hi (x) ≤ 0, (i = 1, . . . , n ) x∈X

(A.29c) (A.29d)

I

where X is the admissible region of x, which also distinguishes the types of the variables, i.e., real, integer, 0–1, etc. The constraints deﬁning the range of variables as xLi ≤ xi ≤ xU (A.30) i , (i = 1, . . . , m) with lower bound xLi and upper bound xU i are called side constraints, bound constraints, or box constraints, and are generally handled separately from the general inequality constraints (A.29c) in the optimization algorithm. A solution is said to be feasible if it satisﬁes all the constraints (A.29b), (A.29c), and (A.29d). The region of feasible solutions is called the feasible region. The feasible solution that minimizes the objective function F (x) is called the global optimal solution or simply the optimal solution. The value of the objective function at the optimal solution is called the optimal objective value or the optimal value. A solution that minimizes F (x) among its neighborhood feasible solutions is called the local optimal solution. The inequality constraint Hi (x) ≤ 0 is said to be active if it is satisﬁed with equality; otherwise, it is inactive.

Appendix

321

An optimization problem is called a linear programming (LP) problem if all the objective and constraint functions are linear functions of real variables. In contrast, the problem is called a nonlinear programming (NLP) problem if at least one of the functions is nonlinear. A solution method for NLP is called nonlinear programming. Contrary to LP, which has general solution methods called the simplex method and the interior-point method, there is no general method that is eﬀective for all types of NLP problems. Therefore, good knowledge is needed to select the most appropriate algorithm for the problem at hand. Since we mainly utilize NLP in this book, details of LP may be consulted in textbooks, e.g., Mangasarian (1969), Peressini, Sullivan, and Uhl (1988), and Luenberger (2003).

A.2.2 A.2.2.1

Nonlinear programming Unconstrained optimization problem

An NLP problem without constraint is called an unconstrained NLP problem. We ﬁrst deﬁne the gradient ∇F (x) of the objective function F (x) with respect to the variables x = (x1 , . . . , xm )⊤ as ( ∇F (x) =

∂F ∂F ,..., ∂x1 ∂xm

)⊤ (A.31)

Note that ∇F (x) corresponds to the direction of maximum increment of F (x) for a speciﬁed value of ∆x⊤ ∆x for the increment ∆x of x. Therefore, the objective function F (x) is most eﬃciently decreased if the solution is modiﬁed in the direction −∇F (x). The optimization method of updating x as follows in the direction −∇F (x) is called the steepest descent method: x(k+1) = x(k) − τ ∇F (x(k) )

(A.32)

where k is the iteration counter, and τ is the parameter deﬁning the magnitude of modiﬁcation of x. In this approach, however, the estimated point overshoots the minimum point in the direction −∇F (x) if τ is too large. In contrast, the computational cost increases if a smaller value is given for τ . Therefore, only the direction is deﬁned by −∇F (x), and the minimum point in this direction is found by the process called line search. Example A.2 Consider the following function deﬁned with respect to two variables x1 and x2 (Katoh, Ohsaki, and Tani 2002): F (x) =

1 (x1 − 3)2 + (x2 − 2)2 2

(A.33)

322

Optimization of Finite Dimensional Structures

x2

∇F(x)

4 3

A

2

C

B

1 1

2

3

4

5

6

x1

FIGURE A.4: Contour lines and gradient of a two-variable function. The ellipses in Fig. A.4 indicate the contour lines of F (x). The gradient of F (x) is given as ( ) x1 − 3 ∇F (x) = (A.34) 2(x2 − 2) For example, at point ‘A’ with (x1 , x2 ) = (4, 3), F (x) = 3/2 and ∇F (x) = (1, 2)⊤ , as shown in Fig. A.4. Suppose point ‘A’ in Fig. A.4 is given as the initial solution x(0) . Then the variables are modiﬁed in the direction −∇F (x(0) ) = (−1, −2)⊤ . Because the objective function is a quadratic function of x, the value of F along the line x = x(0) − τ ∇F (x(0) ) is a quadratic function of τ as F =

1 (1 − τ )2 + (1 − 2τ )2 2

(A.35)

From the stationary condition of F in (A.35) with respect to τ , we obtain τ = 5/9, from which x(1) = (31/9, 17/9)⊤ is derived. Hence, x is updated to x(1) , indicated by point ‘B’ in Fig. A.4, where the gradient is ∇F (x) = (4/9, −2/9)⊤ . By carrying out the line search again, we move to x(2) = (83/27, 56/27), which is indicated by point ‘C’. This way, the point close to the optimal solution (3, 2) has been obtained with two steps of iteration for this small example with ellipsoidal contour lines of F (x). However, if the contour lines have irregular shapes, convergence is very slow; furthermore, a local optimal solution may be found for a nonconvex objective function. Since the objective function is not usually a quadratic function, a more general approach is needed for the line search. Among several algorithms, including equal-interval search and golden section search, the simplest approach is the bi-section search as follows: Step 1: Suppose the solution x(k) is obtained at the kth iteration of the steepest descent method. Initialize the parameter as τ [0] = τ0 with the

Appendix

323

speciﬁed value τ0 , and set the iteration counter of line search j = 0 and x[0] = x(k) . Step 2: Evaluate F (x[j+1] ) at x[j+1] = x[j] − τ [j] ∇F (x(k) ). Step 3: If |F (x[j+1] ) − F (x[j] )| is smaller than the speciﬁed value, then terminate the line search, let x(k+1) ← x[j] , and go to the next step of the steepest descent method. Step 4: If F (x[j+1] ) < F (x[j] ), let τ [j+1] = τ [j] ; otherwise let τ [j+1] = −τ [j] /2. Update the iteration counter as j ← j + 1 and go to Step 2. The optimality of the solution can be veriﬁed using the stationary condition ∇F (x) = 0

(A.36)

and the positive semideﬁniteness of the Hessian D, for which the (i, j)-component Dij is deﬁned as ∂2F (A.37) Dij = ∂xi ∂xj A.2.2.2

Constrained optimization problem: equality constraints

In structural design problems in various ﬁelds of engineering, design requirements are usually given in the form of constraints on responses, such as stresses and displacements under a speciﬁed design load. Since the responses are generally nonlinear functions of the design variables, constrained nonlinear programming problems are to be solved for optimizing structures. For a problem with equality constraints only, the conditions for optimality are obtained using the Lagrange multiplier approach (Bersekas 1982). Let λj denote the Lagrange multiplier for the equality constraint Gj (x) = 0, and deﬁne the Lagrangian as n ∑ E

ψ(x, λ) = F (x) +

λj Gj (x)

(A.38)

j=1

where λ = (λ1 , . . . , λnE )⊤ . In the following, the argument x is omitted for brevity. By diﬀerentiating ψ(x, λ) with respect to xi , we obtain the following stationary conditions: ∑ ∂Gj ∂F ∂ψ = + = 0, (i = 1, . . . , m) λj ∂xi ∂xi j=1 ∂xi nE

(A.39)

The unknown variables x1 , . . . , xm and multipliers λ1 , . . . , λnE are obtained from nE constraints Gj (x) = 0 and m stationary conditions (A.39). Note

324

Optimization of Finite Dimensional Structures

x2

∆

G1

Optimal solution G 1 = constant F

D

∆

F

G1= 0

x1

O

FIGURE A.5: Gradients of objective and constraint functions at the optimal the solution with a single equality constraint. that the stationary conditions are necessary conditions for optimality, and global optimal solutions cannot always be obtained by the Lagrange multiplier method; i.e., the set of nE + m equations may have multiple solutions. Eq. (A.39) can be written in a vector form as n ∑ E

∇ψ = ∇F +

λj ∇Gj = 0

(A.40)

j=1

It is seen from (A.40) that ∇F is expressed as a linear combination of the gradients ∇Gj of the constraints at the optimal solution: n ∑ E

∇F = −

λj ∇Gj

(A.41)

j=1

As a simple case, if there exists only one constraint G1 = 0 with respect to two variables, ∇F and ∇G1 are in the same or the opposite direction, as illustrated in Fig. A.5, and the contour line of the constraint is tangential to that of the objective function at the optimal solution. Example A.3 Consider the following quadratic programming (QP) problem (Katoh, Ohsaki, and Tani 2002): 1 (x1 − 3)2 + (x2 − 2)2 2 subject to G1 (x1 , x2 ) = x1 + 3x2 − 12 = 0

Minimize

F (x1 , x2 ) =

(A.42a) (A.42b)

Appendix

325

x2 ∇F(x) Optimal solution (39/11, 31/11)

4 3

G1 (x) = 0

2 1

∇G1 (x) 1

2

3

4

5

6

x1

FIGURE A.6: Optimal solution of an equality constrained nonlinear programming problem. The Lagrangian is given as ψ(x1 , x2 , λ1 ) =

1 (x1 − 3)2 + (x2 − 2)2 + λ1 (x1 + 3x2 − 12) 2

(A.43)

The stationary conditions of the Lagrangian are obtained as x1 − 3 + λ1 = 0, 2x2 − 4 + 3λ1 = 0

(A.44)

from which and constraint (A.42b), we have the following optimal solution: x1 =

31 6 39 , x2 = , λ=− 11 11 11

In fact, the following relations hold at the optimal solution: ( ) ( ) 6 1 1 , ∇G1 = , ∇F = −λ1 ∇G1 ∇F = 3 11 3

(A.45)

(A.46)

which is illustrated in Fig. A.6. This way, a QP problem with a convex quadratic objective function and linear equality constraints can be solved explicitly using the Lagrange multiplier approach. A.2.2.3

Constrained optimization problem: inequality constraint

Consider an NLP problem with inequality constraints Hi (x) ≤ 0 (i = 1, . . . nI ), where equality constraints are not considered for brevity. As an example, the feasible region is illustrated in Fig. A.7 for a problem with two variables x1 , x2 , and two constraints H1 (x) ≤ 0, H2 (x) ≤ 0. The Lagrangian is deﬁned as nI ∑ ψ(x, µ) = F (x) + µj Hj (x) (A.47) j=1

326

Optimization of Finite Dimensional Structures

x2 H1 H2

∆

∆ feasible region

H1= 0

H2 = 0

∆

F

x1

O

FIGURE A.7: Gradients of objective and constraint functions at the optimal solution with inequality constraints. where µj (≥ 0) is the Lagrange multiplier for the inequality constraint Hj (x) ≤ 0. Roughly speaking, a solution is locally optimal if there is no feasible neighborhood solution that has a smaller objective value than the current solution. Therefore, F should not decrease in the direction in which all of the functions of the active constraints satisfying Hj = 0 decrease. On the other hand, the inactive constraint satisfying Hj < 0 does not have any eﬀect on the optimality of the solution. Hence, as illustrated in Fig. A.7, the gradient ∇F of the objective function should be expressed as a non-positive linear combination of the gradients of the active constraints as n ∑ I

∇F = −

µj ∇Hj

(A.48)

j=1

with

{

µj ≥ 0 µj = 0

for Hj = 0 for Hj < 0

(A.49)

Eq. (A.48) is rewritten as n ∑ I

∇ψ = ∇F +

µj ∇Hj = 0

(A.50)

j=1

Eqs. (A.49) and (A.50) together with constraints Hi ≤ 0 are called optimality conditions or Karush-Kuhn-Tucker (KKT) conditions, and (A.49) is called complementarity conditions. The solution satisfying the KKT conditions is called the KKT point in the variable space. Note that the KKT conditions are the ﬁrst-order necessary conditions for local optimality.

Appendix

327

Suppose the variable vector is modiﬁed by ∆x from the solution satisfying the KKT conditions. Then, using the linear approximation ∆F = ∇F ⊤ ∆x, ∆Hj = ∇Hj⊤ ∆x

(A.51)

the increment of the objective function is approximated with (A.48) as n ∑ I

∆F = −

µj ∆Hj

(A.52)

j=1

Therefore, ∆F ≥ 0 is satisﬁed with ﬁrst-order approximation for any increment ∆x in the direction of the feasible region satisfying ∆Hj ≤ 0 for the active constraints. The KKT conditions are suﬃcient conditions for global (local) optimality if all the objective and constraint functions are globally (locally) convex; i.e., the solution satisfying the KKT conditions is locally optimal if the Hessian of the Lagrangian is positive semideﬁnite. More general and rigorous suﬃcient conditions and second-order optimality conditions may be consulted in the textbooks, e.g., Floudas (1995). If we have bound constraints xLi ≤ xi ≤ xU i , the Lagrangian is reformulated as n ∑ I

ψ(x, µ, η, κ) = F (x) +

µj Hj (x) +

j=1

m ∑

ηi (xi − xU i )+

i=1

m ∑

κi (xLi − xi ) (A.53)

i=1

where ηi and κi are the nonnegative Lagrange multipliers for the bound constraints. Then, from the stationary conditions and complementarity conditions, we have the following optimality conditions: Zi = 0 for xLi < xi < xU i Zi ≤ 0 for xi = xU , (i = 1, . . . , m) (A.54) i Zi ≥ 0 for xi = xLi with the complementarity conditions (A.49), where ∑ ∂Hj ∂F Zi = + , (i = 1, . . . , m) µj ∂xi j=1 ∂xi nI

(A.55)

For a problem with side constraints only, Zi in (A.54) turns out to be the derivative of the objective function F (x). The optimality conditions can be directly solved to obtain the optimal solution (see, e.g., Sec. 1.7 for the optimality criteria approach for structural optimization). The optimality conditions are also used as the termination conditions for a gradient-based nonlinear programming approach.

328

A.2.2.4

Optimization of Finite Dimensional Structures

Penalty function approach

The most convenient approach for incorporating the constraints into the optimization algorithm is the penalty function approach, or penalty approach, which simply increases the objective function value by adding the term of penalization for the violated constraints; thus the constrained problem is transformed into an unconstrained problem. This method is widely used for heuristic methods, e.g., genetic algorithms; see Sec. 6.6. Consider, for simplicity, a minimization problem of an objective function under inequality constraints only: Minimize

F (x)

(A.56a)

subject to Hi (x) ≤ 0, (i = 1, . . . , n ) I

(A.56b)

A penalty function R(Hj (x)) (> 0) is given for the inequality constraint Hj ≤ 0. Let r denote the positive penalty parameter. Then the objective function F (x) is transformed to F ∗ (x, r) as n ∑ I

∗

F (x, r) = F (x) + r

R(Hj (x))

(A.57)

j=1

and the following unconstrained problem is to be solved: Minimize F ∗ (x, r)

(A.58)

where the parameter r is ﬁxed while optimizing with respect to the variable vector x. Penalty function approaches are classiﬁed into the exterior penalty approach and the interior penalty approach depending on the deﬁnition of the penalty function. There exist several deﬁnitions of exterior penalty functions, among which the simplest one is the quadratic function: { R(Hj (x)) = (Hj (x))2 for Hj (x) > 0 (A.59) R(Hj (x)) = 0 for Hj (x) ≤ 0 Suppose the optimal solution of the original constrained problem (A.56) exists at the boundary of the feasible region deﬁned by the inequality constraints; i.e., at least one inequality constraint is satisﬁed with equality at the optimal solution. Then, in the exterior penalty approach, the constraint is slightly violated at the optimal solution of Problem (A.56), and the amount of violation is large if r is small. Therefore, if the violation is larger than the speciﬁed tolerance, r is increased and an approximate solution is searched again by solving the unconstrained problem (A.58). Hence, a very large value should be assigned for r to obtain the solution with good accuracy. However, the Hessian of the problem becomes ill-conditioned, and the gradient of F ∗ (x, r) diverges if r is too large. Therefore, a method such as the sequential unconstrained minimization technique (SUMT) (Fiacco and Cormic 1968) can be used for gradually and systematically updating the parameter r.

Appendix

329

Objective function

5 4 3 2 1

Feasible region

0 0

1

2

3

4

5

6

Variable x

FIGURE A.8: Exterior penalty function; solid line: F (x), dotted line: r = 20, dashed line: r = 1. Example A.4 Fig. A.8 illustrates the problem of minimizing the objective function F (x) =

1 x+1 3

(A.60)

under simple bound constraints 2≤x≤4

(A.61)

The inequalities in (A.61) are converted to standard forms as H1 (x) = −x + 2 ≤ 0

(A.62a)

H2 (x) = x − 4 ≤ 0

(A.62b)

Hence, the penalized objective function is formulated as F ∗ (x, r) =

1 x + 1 + r(R(H1 ) + R(H2 )) 3

(A.63)

Suppose x < 2, i.e., the constraint H1 (x) ≤ 0 is not satisﬁed. Then (A.63) is written as 1 F ∗ (x, r) = x + 1 + r(−x + 2)2 (A.64) 3 From (A.64), the stationary condition of F ∗ with respect to x is obtained as 1 dF ∗ = 2rx − 4r + = 0 dx 3

(A.65)

Therefore, the approximate optimal solution x e that minimizes F ∗ is explicitly written as 1 (A.66) x e=2− 6r

Optimization of Finite Dimensional Structures

330

Objective function

5 4 3 2 1

Feasible region

0 0

1

2

3

4

5

6

Variable x

FIGURE A.9: Interior penalty function; solid line: F (x), dotted line: r = 0.6, dashed line: r = 0.1. Fig. A.8 shows the plots of F ∗ (x, r) corresponding to two diﬀerent values of r as well as the original objective function. As is seen from (A.66), the error 1/(6r) decreases as r is increased. Therefore, x e approaches the optimal solution x = 2 as r is increased. However, if r is too large, the diﬀerential coeﬃcient dF ∗ /dx diverges, and the convergence property of the optimization process for minimizing F ∗ will deteriorate if a gradient-based optimization algorithm is used. Another disadvantage of this approach is that the approximate optimal solution always exists outside the feasible region if there exist active constraints at the optimal solution. In the interior penalty function approach, the following logarithmic barrier function is usually used: n ∑ I

∗

F (x, r) = F (x) − r

log(−Hj (x))

(A.67)

j=1

where the penalty parameter r should be suﬃciently small. Fig. A.9 illustrates the penalized objective functions for the problem deﬁned by (A.60) and (A.61). As is seen, a smaller value of r leads to a better approximate optimal solution. The main advantage of this approach is that the solution always exists in the feasible region. However, this is also a disadvantage because log(−Hj (x)) is not deﬁned for Hj (x) > 0; hence, only the feasible solutions should be searched, which may be very diﬃcult for a large-scale optimization problem. There are various types of interior penalty function approaches, which are simply called the interior point method, for quadratic programming (QP) and semideﬁnite programming (SDP); see Appendix A.2.4 for a brief introduction to SDP. For these problems, the solution of minimizing the penalized objective

Appendix

331

function in (A.67) is traced parametrically with respect to r by solving the KKT conditions; see, e.g., Kojima, Shindoh, and Hara (1997) and Mehrotra (1992) for details. A.2.2.5

Sequential quadratic programming

Sequential quadratic programming (SQP) is an extension of sequential linear programming (SLP), which solves linearized subproblems successively using the standard approaches of LP to ﬁnd the optimal solution of a general NLP problem (Vanderplaats 1999; Gill, Murray, and Saunders 2002; Arora 2004). Taylor’s expansion of the objective function F (x) as well as the constraint functions Gj (x) and Hj (x) at the solution x(k) at the kth step of solving the original NLP problem (A.29) leads to F (x) = F (x(k) ) + ∇F (x(k) )⊤ (x − x(k) ) + · · · Gj (x) = Gj (x

(k)

Hj (x) = Hj (x

(k) ⊤

(k)

(k) ⊤

(k)

) + ∇Gj (x

(k)

) + ∇Hj (x

) (x − x

) (x − x

(A.68a)

) + ···

(A.68b)

) + ···

(A.68c)

If we use only linear terms and deﬁne ∆x(k) = x(k+1) − x(k) , the following LP subproblem of SLP is derived: Minimize

F (x) = F (x(k) ) + ∇F (x(k) )⊤ ∆x(k) (k) ⊤

(k)

= −Gj (x

(k) ⊤

(k)

≤ −Hj (x

subject to ∇Gj (x

∇Hj (x

) ∆x

) ∆x

(k)

(A.69a) E

(A.69b)

I

(A.69c)

), (i = 1, . . . n )

(k)

), (i = 1, . . . n )

which is solved for the variable vector ∆x(k) = (∆x1 , . . . , ∆xm )⊤ using the method of LP. Then, x is updated to x(k+1) = x(k) + ∆x(k) , and subproblem (A.69) is solved successively. (k)

(k)

However, in SLP, the error due to linearization may be very large if there is no upper bound for the magnitude of ∆x(k) . Therefore, we usually assign (k) (k) the bound ∆xi , called the move limit for the absolute value |∆xi | of each (k) component of ∆x , which is proportionally reduced by a factor c (< 1) as (k+1) (k) = c∆xi . The algorithm converges if the norm of ∆x(k) becomes ∆xi suﬃciently small. However, there will be no feasible solution if the current solution is infeasible and the move limit is too strict. Furthermore, the solution converges to a non-optimal solution if the move limit is reduced too quickly. Hence, it is very diﬃcult to ﬁnd an appropriate value of the reduction factor c. To alleviate this diﬃculty, SLP is extended to SQP, where a quadratic penalty is given for the magnitude of ∆x(k) as follows to formulate a quadratic

332

Optimization of Finite Dimensional Structures

programming subproblem: 1 F (x) = F (x(k) ) + ∇F (x(k) )⊤ ∆x(k) + ∆x(k)⊤ D∆x(k) (A.70a) 2 subject to ∇Gj (x(k) )⊤ ∆x(k) = −Gj (x(k) ), (i = 1, . . . nE ) (A.70b) Minimize

∇Hj (x(k) )⊤ ∆x(k) ≤ −Hj (x(k) ), (i = 1, . . . nI )

(A.70c)

Although D may be any m × m positive deﬁnite matrix in view of assigning a penalty for the large magnitude of solution update ∆x(k) , an approximate Hessian of the Lagrangian of the original problem is preferred in order to assure convergence to a local optimum. The Hessian or its inverse can be appropriately computed using an algorithm of the quasi-Newton method. After solving Problem (A.70) for ∆x(k) , the solution may be simply updated to x(k) + ∆x(k) ; however, the best solution is usually searched in the direction of ∆x(k) using the technique of line search in Sec. A.2.2.1 to minimize a descent function or merit function deﬁned by the objective function and the penalty for violation of the constraints. In this book, we mainly use the software SNOPT Ver. 7.2 (Gill, Murray, and Saunders 2002), which consists of a major iteration which utilizes a quasi-Newton method called the Broyden–Fletcher–Goldfarb–Shanno (BFGS) method for approximately updating the Hessian, and a minor iteration for solving the QP subproblem using a reduced-gradient method. An augmented Lagrangian merit function is used for computing the step size of the line search. A.2.2.6

Method of feasible directions

The simplest approach to an unconstrained NLP problem is to update the solution in the steepest descent direction of the objective function, as presented in Sec. A.2.2.1. For a constrained problem, the gradient of the objective function can be projected to the hyper-plane of the active constraints to obtain a feasible direction of solution update. However, because the constraints are generally nonlinear, the updated solution very often violates the constraints, and we cannot allow a large magnitude of solution update using this approach. In the method of feasible directions (MFD), the search direction S(k) at the kth step is directed into the feasible region, and the solution is updated as x(k+1) = x(k) + αS(k)

(A.71)

where the scaling parameter α is found by line search (Ben-Israel, Ben-Tal, and Zolbec 1981; Vanderplaats 1999). Suppose we have only inequality constraints, and let J denote the set of active constraints at the kth iteration. Then the usable direction S(k) is char-

Appendix

333

acterized by ∇F (x(k) )⊤ S(k) ≤ 0,

(A.72)

∇Hj (x(k) )⊤ S(k) ≤ 0, (j ∈ J )

which means that S(k) is directed into the inside of the feasible region and the objective function decreases as the solution is modiﬁed in the direction of S(k) . However, the best direction that minimizes ∇F (x(k) )⊤ S(k) in the feasible direction may be generally tangent to the boundary of the feasible region. Therefore, a pushoﬀ factor θ is introduced, and the auxiliary variable β is maximized by solving the following LP problem: Minimize

β

(A.73a)

subject to ∇F (x

(k) ⊤ (k)

) S

∇Hj (x

) S

S ≤S L

+β ≤0

(k) ⊤ (k) (k)

≤S

(A.73b)

+ θβ ≤ 0, (j ∈ J )

(A.73c)

U

(A.73d)

where the variables are S(k) and β, and SL and SU are the lower and upper bounds for S(k) . This way, a feasible direction that reduces the objective function is obtained, and the best solution is searched in this direction using a line search. A.2.2.7

Method of moving asymptote

The method of moving asymptote (MMA) developed by Svanberg (1987) is based on a successive convex approximation using the reciprocals of the variables. Therefore, the MMA may be regarded as an extension of the convex linearization method (CONLIN) (Fleury 1989a). Consider an optimization problem with inequality constraints Hi (x) ≤ 0 L(k) U(k) (i = 1, . . . , nI ) and side constraints xL ≤ x ≤ xU . Let Mi and Mi denote the moving asymptotes at the kth iteration given for xi as L(k)

Mi

(k)

< xi

U(k)

< Mi

(A.74) L(k)

U(k)

The variable xi is converted to the reciprocal 1/(xi − Mi ) or 1/(Mi − xi ) in view of the signs of the derivatives of the objective function and the (k) constraint functions with respect to xi . Let ∂Hj /∂xi denote the value of ∂Hj /∂xi at x = x(k) . The approximation Hja (x) of Hj (x) is deﬁned with respect to the reciprocals as Hja (x)

=

(k) rj

+

m ∑

(

U(k)

i=1

Mi

− xi

)

(k)

(k)

pij

+

qij

L(k)

xi − Mi

(A.75)

Optimization of Finite Dimensional Structures

334

where

(k)

pij

(k)

qij

(k)

rj

(k) (k) ∂Hj ∂Hj U(k) 2 for >0 − xi ) (Mi ∂xi ∂xi = (k) ∂Hj 0 ≤0 for ∂xi (k) ∂Hj ≥0 for 0 ∂xi = (k) ∂Hj −(xi − M L(k) )2 ∂Hj for <0 i ∂xi ∂xi ) ( (k) (k) m ∑ qij pij (k) + (k) = Hj (x ) − U(k) (k) L(k) Mi − xi xi − Mi i=1 (k)

(k)

(A.76)

(A.77)

(A.78)

(k)

The objective function is expanded similarly with pi0 , qi0 , and r0 deﬁned by assuming H0 (x) = F (x) in (A.75)–(A.78). Then the following subproblem is to be solved: ( ) m (k) (k) ∑ pi0 qi0 (k) Minimize r0 + + (A.79a) U(k) L(k) Mi − xi xi − Mi i=1 ) ( (k) (k) m ∑ qij pij (k) + ≤ 0, subject to rj + U(k) L(k) Mi − xi xi − Mi i=1 (j = 1, . . . , N I ) αiL ≤ xi −

(k) xi

≤ αiU

(A.79b) (A.79c)

where αiL and αiU are the appropriately assigned move limits satisfying L(k)

Mi

A.2.3

(k)

< αiL + xi

(k)

≤ xi ≤ αiU + xi

U(k)

< Mi

(A.80)

Dual problem

As a simple example, consider ﬁrst an LP problem Minimize c⊤ x subject to Bx − b ≥ 0 x≥0

(A.81a) (A.81b) (A.81c)

where c = (c1 , . . . , cm )⊤ and b = (b1 , . . . , bn )⊤ are constant vectors, and B is an n × m constant matrix.

Appendix

335

The dual problem of the primal problem (A.81) is formulated using the dual variables µ = (µ1 , . . . , µn )⊤ as Maximize b⊤ µ

(A.82a)

⊤

subject to B µ − c ≤ 0 µ≥0

(A.82b) (A.82c)

where µ coincides with the vector of Lagrange multipliers of the constraints (A.81b) of the primal problem. For the feasible solutions x and µ of problems (A.81) and (A.82), respectively, the following inequality holds: c⊤ x ≥ µ⊤ Bx ≥ µ⊤ b

(A.83)

where (A.81b) and (A.82b) have been used. Therefore, the optimal value of Problem (A.81) is not less than that of Problem (A.82). This condition is called weak duality. Furthermore, the equality in (A.83) is satisﬁed by the optimal solutions of (A.81) and (A.82), which is called strong duality. Consider next the following NLP problem with inequality constraints only: Minimize

F (x)

(A.84a)

subject to Hi (x) ≤ 0, (i = 1, . . . , nI )

(A.84b)

The Lagrangian for this problem is given as n ∑ I

ψ(x, µ) = F (x) +

µj Hj (x)

(A.85)

j=1

There are several deﬁnitions of a dual problem for an NLP problem. For example, Wolf’s dual problem is formulated as n ∑ I

Maximize ψ(x, µ) = F (x) +

µj Hj (x)

(A.86a)

j=1 n ∑ I

subject to ∇F (x) +

µj ∇Hj (x) = 0

(A.86b)

j=1

µi ≥ 0, (i = 1, . . . , nI )

(A.86c)

where the variables are x and µ. Let xP and (xD , µ) denote arbitrary feasible solutions of Problems (A.84) and (A.86), respectively. Then the following weak duality holds: F (xP ) ≥ ψ(xD , µ)

(A.87)

336

Optimization of Finite Dimensional Structures

Furthermore, if F (x) and Hi (x) are convex, and the primal problem (A.84) e, then there exists the optimal solution (e e ) of dual has an optimal solution x x, µ problem (A.86) satisfying the strong duality condition: e) F (e x) = ψ(e x, µ

(A.88)

Example A.5 As a small example, consider the following simple convex quadratic programming problem with two variables: Minimize x21 + x22 subject to x1 + x2 + 1 ≤ 0

(A.89a) (A.89b)

The dual problem is formulated as Maximize x21 + x22 + µ1 (x1 + x2 + 1)

(A.90a)

subject to 2x1 + µ1 = 0 2x2 + µ1 = 0 µ1 ≥ 0

(A.90b) (A.90c) (A.90d)

From the constraints, we have x1 = x2 = −µ1 /2, and the objective function is expressed with respect to µ1 only as −µ21 /2 + µ1 . Therefore, the optimal solution of the dual problem is obtained as x1 = x2 = −1/2, µ1 = 1, which leads to the optimal objective value 1/2, which is the same as that of the primal problem. As another formulation, the Lagrange dual problem is deﬁned as Maximize Φ(µ) = min ψ(x, µ)

(A.91a)

subject to µ ≥ 0

(A.91b)

x

For example, consider again the simple primal problem (A.89). The solution that minimizes ψ in the dual problem (A.91) is obtained from the stationary conditions (A.90b) and (A.90c) of ψ, which leads to x1 = x2 = −µ1 /2. Then, the objective function is given as Φ(µ1 ) = −µ21 /2 + µ1 . Hence, maximization of Φ(µ1 ) leads to µ1 = 1 and Φ(µ1 ) = 1/2, which is equal to the optimal value of the primal problem; i.e, the strong duality is satisﬁed.

A.2.4

Semideﬁnite programming

An optimization problem that has constraints such that the variable matrix is positive semideﬁnite is called a semideﬁnite programming (SDP) problem. The objective function and other constraints are linear in the standard form of an SDP problem (Wolkowicz, Saigal, and Vandenberghe 2000; Ohsaki and

Appendix

337

Kanno 2007). Since the constraints of positive semideﬁniteness of matrices include linear and convex quadratic constraints, the SDP is an extension of linear programming and convex quadratic programming. Interior-point methods for linear and quadratic programming have been extended to solve SDPs (Kojima, Shindoh, and Hara 1997). Let X ≽ O indicate that the symmetric matrix X is positive semideﬁnite. The inner product X • Y of the n × n matrices X = (Xij ) and Y = (Yij ) is deﬁned as X•Y =

n ∑ n ∑

Xij Yij

(A.92)

i=1 j=1

The standard form of the SDP problem is given as Minimize C • X subject to Ai • X = bi , (i = 1, . . . , m) X≽O

(A.93a) (A.93b) (A.93c)

where n × n matrices C and Ai , and vector b = (b1 , . . . , bm )⊤ are constant, and X is the variable matrix. The dual problem of Problem (A.93) is formulated as b⊤ y m ∑ subject to Ai y i + Z = C

Minimize

(A.94a) (A.94b)

i=1

Z≽O

(A.94c)

where the vector y = (y1 , . . . , ym )⊤ and n × n matrix Z are the variables. The KKT conditions for the pair of primal and dual SDP problems are written as XZ = O

(A.95)

with the constraints (A.93b), (A.93c), (A.94b), and (A.94c) of the primal and dual problems. The algorithm for solving the KKT conditions incorporating the relaxation parameter for complementary conditions (A.95) and interior penalty function is generally called the primal-dual interior-point method. Various structural optimization problems, including the robust optimization problem (Ben-Tal and Nemirovski 1997) and the truss topology optimization problem under frequency constraint (Ohsaki, Fujisawa, Katoh, and Kanno 1999), can be formulated as an SDP problem; see Sec. 3.9.

338

A.2.5

Optimization of Finite Dimensional Structures

Combinatorial problem

If the variables can take only integer values, then the optimization problem is called an integer programming (IP) problem, which is formulated as Minimize

F (x)

(A.96a) E

subject to Gi (x) = 0, (i = 1, . . . , n )

(A.96b)

Hi (x) ≤ 0, (i = 1, . . . , n )

(A.96c)

x∈Z

(A.96d)

I

m

where Z denotes the set of integers, and m is the number of variables. Since the problem of ﬁnding the optimal combination of the variables is formulated as an IP problem, it is equivalently called a combinatorial optimization problem. If some of the variables can take real values, then the problem is called a mixed integer programming (MIP) problem, which usually denotes a problem with linear objective and constraint functions. An MIP with nonlinear objective and/or constraint functions is called a mixed integer nonlinear programming (MINLP) problem, which often appears in structural optimization problems with discrete design variables and continuous state variables. The most standard approach to IP is the branch-and-bound method (Horst, Pardalos, and Thoai 1995; Floudas 1995), where the original problem is successively divided into subproblems, and the upper and lower bounds of the optimal objective value are updated by solving the relaxed problem assuming that the variables can have real numbers. Note that all the objective and constraint functions of the relaxed problem should be convex in order to ﬁnd the global solution of the original problem. In the process of branch-and-bound, the subproblem that remains to be solved is called the active problem. Since any IP can be reformulated as a 0–1 problem, we present the basic algorithm of IP, as follows, assuming the variables xi can take either 0 or 1, and F (x), Gi (x), and Hi (x) are linear functions of x: Step 0: Initialize the upper bound F U of F (x) as F U = ∞. Let the set A of the active problems consist of the original IP (A.96). Step 1: Select a problem P from A and remove it from A. Note that the number of branching processes can be reduced if a smaller initial value is given for F U using, e.g., a heuristically obtained approximate optimal solution. Step 2: Solve the relaxed LP problem P of P by relaxing the integer conditions of the variables to inequality constraints 0 ≤ xi ≤ 1, and select a variable j satisfying 0 < xj < 1 at the optimal solution of P . Let P0 and P1 denote the subproblems of P by specifying xj = 0 and 1, respectively. Solve the relaxed LPs P 0 and P 1 , respectively, of P0 and P1 .

Appendix

339

Step 3: Let F0 and F1 denote the optimal objective values of P 0 and P 1 , respectively. If F0 > F U , then set xj = 1 and terminate P0 . If F1 > F U , then set xj = 0 and terminate P1 . Step 4: If the solution of P 0 satisﬁes the integer conditions, then let F U = min{F0 , F U }, and terminate P0 ; otherwise, add P0 to A. Step 5: If the solution of P 1 satisﬁes the integer conditions, then let F U = min{F1 , F U }, and terminate P1 ; otherwise, add P1 to A. Step 6: If A ̸= ∅, then go to Step 1; otherwise, output the best value of F U , and terminate the process. Example A.6 As a simple example, consider the following LP problem: P : Minimize x1 + x2 subject to − 2x1 − x2 + 4 ≤ 0

(A.97a) (A.97b)

− x1 − 4x2 + 4 ≤ 0 x1 ∈ Z, x2 ∈ Z

(A.97c) (A.97d)

which is solved using the branch-and-bound method as follows: 1. At the initial stage, let F U = ∞. The set A of active problems consists of P , which is to be selected and solved. 2. The optimal solution of the relaxed LP of P , denoted by P , without the integer constraint (A.97d) is found as (x1 , x2 ) = (12/7, 4/7). 3. We select x1 for generating the two subproblems P0 and P1 of P with x1 = 1 and x1 = 2, respectively, and formulate the following relaxed problems P 0 and P 1 , respectively, of P0 and P1 : P 0 : Minimize 1 + x2 subject to − 2 − x2 + 4 ≤ 0 − 2 − 4x2 + 4 ≤ 0

(A.98a) (A.98b) (A.98c)

x2 ∈ R

(A.98d)

2 + x2

(A.99a)

subject to − 4 − x2 + 4 ≤ 0 − 1 − 4x2 + 4 ≤ 0

(A.99b) (A.99c)

x2 ∈ R

(A.99d)

P 1 : Minimize

where R is the set of real numbers.

340

Optimization of Finite Dimensional Structures

4. The optimal values F0 and F1 of P 0 and P 1 , respectively, are found as 3 and 5/2, and the corresponding optimal solutions are (x1 , x2 ) = (1, 2) and (2, 1/2). 5. Because the solution of P 0 satisﬁes the integer conditions and F0 < F U = ∞, F U is updated as F U = F0 = 3, and P0 is terminated. 6. Because the solution of P1 does not satisfy the integer conditions, P 1 is added to the set A of the active problems. 7. Problem P1 is selected from A to formulate subproblems P10 and P11 by ﬁxing x2 to 0 and 1, respectively. Since no variable is left, and P10 is infeasible, we can immediately compute the objective value of P11 as F1 = 3, which is the same as the current upper bound F U . 8. Since no problem is left in A, optimal solutions are found to be (x1 , x2 ) = (1,2) and (2,1) with the same objective value F (x) = 3.

A.3 A.3.1

Heuristics Introduction

Heuristic approaches (or heuristics for simplicity) have been developed to obtain approximate optimal solutions within reasonable computational cost, although there is no theoretical proof of convergence to the global optimal solution (Reeves 1995). Among many heuristic approaches, the most popular approach is the genetic algorithm (GA) (Goldberg 1989), which can be categorized as a multipoint search or population-based approach that has many solutions at each iterative step called generation. Although a GA generally requires a very large population size, some methods, e.g., micro-GA, scatter search (Laguna and Marti 2003), particle swarm optimization (Kennedy 1997), immune algorithm, and harmony search (Geem, Kim, and Loganathan 2001; Lee and Geem 2004), have been developed to ﬁnd approximate optimal solutions with a relatively small population size. However, because the computational cost for evaluating the objective and/or constraint functions at each step is very large for structural optimization problems, a multipoint strategy may not be appropriate, especially for optimization of a structure with a large number of degrees of freedom. Therefore, single-point-search heuristics, including simulated annealing (SA) (Kirkpatrick, Gelatt, and Vecchi 1983; Aarts and Korst 1989; Cerny 1985), and tabu search (or taboo search, TS) (Glover 1989; Glover and Laguna 1997), may have advantages over the multipoint strategies.

Appendix

A.3.2 A.3.2.1

341

Single-point-search heuristics Basic algorithm

Single-point-search heuristics are based on local search (Aarts and Lenstra 1997), in which the solution is consecutively updated to a neighborhood solution if it improves (‘reduces’ for a minimization problem) the value of the objective function, where the neighborhood solutions are generated by modifying the value of one or several variables to neighboring values. Because it is not always possible to ﬁnd a good approximate optimal solution by simple local searches, heuristic approaches have been proposed to improve the convergence properties. The basic algorithm of a single-point-search heuristic approach can be stated as follows: Step 1: Assign an initial solution. Step 2: Carry out local search to select a candidate solution for the next step from the neighborhood solutions of the current solution. Step 3: Accept or reject the candidate solution in accordance with the criteria deﬁned by the speciﬁc algorithm. Step 4: Go to Step 2 if not converged. For the initial solution, we can assign either speciﬁed or randomly generated values. An optimal solution found by using another approach can be modiﬁed to be used as the initial solution; e.g., the nearest discrete solution from the optimal solution with continuous variables can be used. The neighborhood solutions are the set of solutions that can be reached from the current solution by the speciﬁed operation; e.g., for the truss topology optimization problem, the cross-sectional area of a randomly selected member can be increased or decreased by the speciﬁed value, or the locations of a pair of existing and non-existing members are exchanged, and so on. Let J denote the vector of integer variables. Consider a minimization problem of the objective function F (J) with inequality constraints Hj (J) ≤ 0 (j = 1, . . . , nI ). If inequality constraints cannot be directly handled in the algorithm, the following penalty function is used to evaluate the performance of the solution: n ∑ I

∗

F (J) = F (J) +

αj (max{Hj (J), 0})2

(A.100)

j=1

where αj is a penalty coeﬃcient, and no penalty is given if the constraint Hj ≤ 0 is satisﬁed; see Appendix A.2.2.4 for details. Other penalty function approaches such as dynamic penalty, as well as the augmented Lagrangian approach, can also be used (Lagaros, Papadrakakis, and Kokossalakis 2002).

342

A.3.2.2

Optimization of Finite Dimensional Structures

Greedy/stingy method

Heuristic approaches can be classiﬁed into deterministic and probabilistic approaches. The simplest deterministic approach is the greedy method described as follows for a constrained optimization problem: Step 1: Assign an initial solution that does not satisfy the constraints; e.g., choose the smallest value for all variables for the case in which the constraint functions Hj (J) are decreasing functions of J. Step 2: Move to a neighborhood solution which most eﬃciently improves the objective function and constraints, where a penalty function is used for the deﬁnition of the eﬃciency of the solution. Step 3: Go to Step 2 if one of the constraints is not satisﬁed or the solution cannot be improved. In contrast, an approach that starts with a solution satisfying all the constraints and reduces the objective value consecutively is called the stingy method, described as follows: Step 1: Assign an initial solution that satisﬁes all the constraints; e.g., choose the largest value for all variables for the case in which the constraint functions Hj (J) are decreasing functions of J. Step 2: Move to a neighborhood solution which most eﬃciently reduces the objective function. Step 3: Go to Step 2 if one of the constraints is not satisﬁed. Applications of greedy and stingy methods can be found in Secs. 1.12, 5.1, and 6.3. A.3.2.3

Tabu search

The convergence property of the global optimal solution may be enhanced if many solutions are searched before moving to a neighborhood solution, or preferably, all neighborhood solutions may be searched to select the best neighborhood solution. A neighborhood solution that does not reduce (for a minimization problem) the objective value can also be selected to improve the possibility of reaching the global optimal solution. However, in this case, a so-called cycling or loop can occur where a set of neighboring solutions is chosen iteratively. TS has been developed to prevent cycling utilizing the tabu list containing the prohibited solutions that have already been searched. The algorithm of TS is summarized follows, where the superscript ( · )(k) denotes a value at the kth iteration: Step 1: Assign an initial solution J(0) , and initialize the tabu list T to be empty. Set the iteration counter k = 0.

Appendix

343

N (k) Step 2: Generate neighborhood solutions JN and i (j = 1, . . . , n ) of J ∗ move to the best solution J among them that is not included in the tabu list T .

Step 3: Add J∗ to T . Step 4: Remove the oldest solution in T if the length of the list exceeds the speciﬁed value. Step 5: Let J(k+1) = J∗ and k ← k + 1. Go to Step 2 if the termination condition is not satisﬁed; otherwise, output the best solution satisfying the constraints, and terminate the process. TS is conceived as a deterministic approach if the neighborhood solutions are generated in a deterministic manner. Some attributes, including recency, frequency, quality, and inﬂuence of the solution, instead of the solution itself, can be stored in the tabu list (Glover and Laguna 1997). For example, the value or the move of a speciﬁc variable, or a set of variables, can be an attribute. A.3.2.4

Random search

The simplest probabilistic approach is the random search, which works as follows: Step 1: Randomly generate the initial solution. Step 2: Move to the randomly generated neighborhood solution if it satisﬁes the constraints. Step 3: Go to Step 2 if the termination condition is not satisﬁed. A penalty function approach can also be used for a random search for a constrained problem. Although a random search is not eﬃcient in view of the convergence property to the global optimal solution, it is simple and easy to implement. An application of random search to truss topology optimization is presented in Sec. 3.8. A controlled random search was proposed by Price (1983) for improving convergence properties, and was extended to incorporate the concept of simulated annealing by Mohan and Nguyen (1999), and to incorporate local mutation by Kaelo and Ali (2006). A random move can be incorporated into TS for a single objective problem (Hu 1992) and a multiobjective problem (Baykasoglu, Owen, and Gindy 1999b). An algorithm with jump or restart was presented by Li, Priemer, and Cheng (2004) for problems with real variables. The termination (stopping) rules for random search and local search were investigated by Hart (1998). Surrogate models or metamodels, e.g., response surface approximation (Myers and Montgomery 1995) and kriging (Lee and Jung 2007), can be used for reducing the computational cost (Brigham and Aquino 2007).

Optimization of Finite Dimensional Structures

344

A.3.2.5

Simulated annealing

Simulated annealing (SA) was developed to prevent convergence to a local optimal solution by allowing a move to a solution that does not improve the objective function, where the probability of accepting such a solution is deﬁned by the amount of increase (for a minimization problem) of the objective function. The term simulated annealing comes from the fact that it simulates the behavior of the metals in an annealing process. The basic algorithm is as follows: Step 1: Randomly generate the initial solution J(0) , and set the temperature parameter T to a speciﬁed initial value T0 . Assign the range parameter δ for scaling the objective function. Set the iteration counter k = 0. Step 2: Randomly generate the neighborhood solution J∗ of J(k) . Step 3: Deﬁne ∆F as ∆F = F (J∗ ) − F (J(k) )

(A.101)

If ∆F < 0, or if the random number 0 ≤ R < 1 is smaller than P deﬁned by the following equation, accept J∗ and let J(k+1) = J∗ . ( ) ∆F P = exp − (A.102) δTk If J∗ is not accepted, let J(k+1) = J(k) . Step 4: Decrease Tk to Tk+1 on the basis of the speciﬁed rule. Step 5: If the termination condition is not satisﬁed, set k ← k + 1 and go to Step 2; otherwise, output the best solution satisfying the constraints, and terminate the process. The initial temperature is given so that almost all neighborhood solutions are accepted. The termination condition is deﬁned using the acceptance ratio; e.g., (a) the acceptance ratio among the last several steps is less than the speciﬁed value, or (b) the solution does not change within the speciﬁed number of cycles. The temperature can be updated in Step 4 using the Metropolis rule: Tk+1 = cTk (A.103) with the speciﬁed constant c that is slightly less than 1, e.g., 0.95 (Metropolis, Rosenbluth, Rosenbluth, Teller, and Teller 1953). The solutions are usually updated several times at the same temperature to ensure convergence. Alternatively, Tk may be deﬁned as Tk =

K −k T0 K

(A.104)

Appendix

345

where K is a parameter typically ranging from 15 to 30. Originally, for simulating the annealing process of a metal, the so-called Bolzman constant b was used to replace Tk with bTk ; however, b can be incorporated implicitly in the scaling parameter δ of the objective function and/or the deﬁnition of Tk itself. Furthermore, our purpose is to solve an optimization problem and not to simulate the physical process of annealing. Therefore, we do not use the Boltzmann constant in this book.

A.4 A.4.1

Multiobjective programming Deﬁnition of multiobjective programming

So far, in the Appendix, we considered optimization problems with a single objective function. However, in the process of structural design, it is natural to consider multiple objective functions to be minimized or maximized. An optimization problem with multiple objective functions is called a multiobjective programming (MOP) problem, and its solution method is called multiobjective programming (Cohon 1978; Stadler 1979, 1988; Marler and Arora 2004), which is also called multicriteria optimization and vector optimization. The MOP problem of minimizing the nF objective functions F(x) = (F1 (x), . . . , FnF (x))⊤ is formulated as Minimize

F(x)

(A.105a)

subject to Hi (x) ≤ 0, (i = 1, . . . , n ) I

(A.105b)

where only inequality constraints are considered for brevity. A solution that minimizes all the objective functions simultaneously is called an absolutely optimal solution, which does not exist in general. An example of an absolutely optimal solution for nF = 2 is illustrated in Fig. A.10(a) in the objective function space, for which the coordinates are deﬁned by the objective functions. If the absolutely optimal solution does not exist, it is natural to consider a situation in which at least one objective value increases if one of the remaining objective values is decreased. For two feasible solutions x1 and x2 satisfying the constraints (A.105b), if Fi (x1 ) ≤ Fi (x2 ) for i = 1, . . . , nF and Fj (x1 ) < Fj (x2 ) for one of j ∈ {1, . . . , nF }, then x2 is said to be dominated by x1 . If there is no solution that dominates a feasible solution x∗ , then x∗ is called the nondominated solution, the noninferior solution, the compromise solution, or the Pareto optimal solution, which may be simply called Pareto solution. For an MOP problem with continuous variables, generally there exists an inﬁnite number of Pareto optimal solutions, which form a set called the Pareto optimal set or the Pareto front. The most preferred solution is chosen from

346

Optimization of Finite Dimensional Structures

F2

Pareto optimal set

F2

∆ F2 Absolutely optimal solution O

F1

(a) Absolutely optimal solution

∆ F1 O

F1

(b) Trade-oﬀ ratio of Pareto solutions

FIGURE A.10: Feasible regions and optimal solutions in the objective function space. the Pareto set in view of an additional measure of preference. Fig. A.10(b) shows an example of Pareto optimal solutions for nF = 2. The ratio ∆F2 /∆F1 between the two neighboring Pareto solutions deﬁnes the trade-oﬀ ratio of F2 to F1 , which serves as one of the a posteriori measures of preference. The process of selecting the most preferred solution is summarized as follows (Marler and Arora 2004): 1. Approach with a priori information The upper bounds, weight coeﬃcients, or ideal values of objective functions are speciﬁed, and the single-objective problem is solved using the methods of constraint approach, linear weighted sum approach, goal programming, and so on, which are described in Secs. A.4.2, A.4.3, and A.4.4, respectively. The process is terminated if a single Pareto optimal solution is obtained. 2. Approach without a priori information (a) Enumeration of Pareto optimal solutions: Many Pareto optimal solutions are ﬁrst generated or preferably all the Pareto solutions are enumerated for a problem with integer variables. Then the most preferred solution is selected from the Pareto optimal set in view of additional (a posteriori) information. (b) Interactive approach: A tentative information of preference, e.g., a set of weight coeﬃcients and location of the ideal point in the objective function space, is ﬁrst speciﬁed, and the single Pareto optimal solution is obtained by a similar manner as the approach with a priori information. The preference information is next conceived as a parameter that is to be modiﬁed interactively to obtain a more preferred solution.

Appendix

347

Note that the interactive approach should satisfy the following properties: 1. The relation between the variation of the parameter values and the resulting Pareto optimal solution should be clearly correlated; i.e., the Pareto solution obtained after modiﬁcation of the parameter should be estimated a priori by the decision maker. 2. Pareto optimal solutions can be continuously traced in the objective function space for problems with continuous variables and diﬀerentiable functions. The details of these approaches are explained below.

A.4.2

Constraint approach

Suppose the objective functions are numbered in ascending order with respect to importance; i.e., minimizing F1 (x) is less important than minimizing F2 (x), . . . , FnF (x). Then a single-objective optimization problem is formulated, as follows, to minimize F1 (x) by assigning upper bounds F 2 , . . . , F nF , respectively, for F2 (x), . . . , FnF (x). Minimize

F1 (x)

(A.106a)

subject to Fi (x) ≤ F i , (i = 2, . . . , n ) F

Hi (x) ≤ 0, (i = 1, . . . , n ) I

(A.106b) (A.106c)

This approach is called the constraint approach or the ε-constraint approach. Note that the values of the important objective functions are restricted by assigning their upper bounds, whereas the least important objective function is to be minimized; i.e., it is allowed to have an unexpectedly large value. Fig. A.11(a) illustrates the constraint approach for nF = 2, where the gray area is the feasible region. The upper bound can be iteratively modiﬁed to ﬁnd the most preferred solution in an interactive manner. However, the objective value F1 (x) cannot be minimized, if F i (i = 2, . . . , nF ) are too small and there exists no feasible solution. Therefore, the following procedure is usually used: Step 1: Assign the objective functions F1 (x), . . . , FnF (x) in ascending order of importance. Step 2: Solve the single-objective optimization problem for minimizing FnF under constraints (A.106c) to ﬁnd only the optimal objective value FbnF . Set the iteration counter k = 1. Step 3: Assign a moderately small positive allowable value εnF −k+1 for relaxing the value of FnF −k+1 , and minimize FnF −k (x) under constraints (A.106c) and additional constraints Fj (x) ≤ Fbj + εj , (j = nF − k + 1, . . . , nF ) to ﬁnd the optimal objective value FbnF −k .

(A.107)

Optimization of Finite Dimensional Structures

348

F2

F2

(w1 , w2 )

F2 Pareto optimal solution

Pareto optimal solution

O

O

F1

F1 (b) Weighted sum approach

(a) Constraint approach Pareto optimal set

A

F2

F2

C p r Pareto optimal solution

q

B

Ideal point F O

F1

O

*

s

(c) Relation between Pareto solutions and weight coeﬃcients

F1

(d) Goal programming

FIGURE A.11: Conversion of a multiobjective programming problem to single-objective problem. Step 4: Stop if enough objective functions are minimized; otherwise let k ← k + 1 and go to Step 3.

A.4.3

Linear weighted sum approach

In the linear weighted sum approach, the objective functions Fi (X) are scalarized using the weight coeﬃcients wi (> 0) in view of the importance of each objective function, and the following optimization problem is solved: n ∑ F

Minimize

F (x) =

wi Fi (x)

(A.108a)

i=1

subject to Hi (x) ≤ 0, (i = 1, . . . , nI )

(A.108b)

The Pareto optimal solutions are found for each speciﬁed set of wi . This approach is most convenient if appropriate values can easily be determined for

Appendix

349

wi ; however, this is generally diﬃcult, because the ranges, units (dimensions), etc., of the objective functions are diﬀerent. Consider the simple case of nF = 2, and let R = w1 /w2 . Then the slope of the line with constant F (x) at the Pareto solution in the objective function space is −R, as illustrated in Fig. A.11(b). Therefore, various Pareto optimal solutions are found by modifying R. However, if the feasible region in the objective function space is not convex, Pareto solutions are found discontinuously even if R is modiﬁed continuously, and the complete set of Pareto solutions may not be found using this method. For example, for the case illustrated in Fig. A.11(c), the line with constant F (x) is indicated by ‘A’, if R is moderately large, and the Pareto solution ‘p’ is found. On the other hand, if R is moderately small, the line with constant F (x) is indicated by ‘B’, and the Pareto optimal solution ‘q’ is found. Suppose the line with constant F (x) is given by line ‘C’, and F (x) has the same value at solutions ‘r’ and ‘s’. Then, a solution near ‘r’ is found if R is slightly increased, and a solution near ‘s’ is found if R is slightly decreased. Therefore, the Pareto solutions vary discontinuously with respect to the variation of R; i.e., the solutions between ‘r’ and ‘s’ cannot be found by this approach. Hence, the linear weighted sum approach is not suitable for interactive search of the most preferred solution, because the decision maker cannot expect a change in the solution as the result of modiﬁcation of the weight coeﬃcients. There exist similar approaches, for instance, minimization of the product of F1 (x), . . . , FnF (x), which corresponds to minimization of the sum of log[Fi (x)]. Alternatively, the maximum value of F1 (x), . . . , FnF (x) may be minimized.

A.4.4

Goal programming

The approach for minimizing the distance to the ideal point F∗ in the objective function space is called goal programming, which is formulated as D(F∗ , F(x))

Minimize

(A.109a)

subject to Hi (x) ≤ 0, (i = 1, . . . , n ) I

(A.109b)

where D(F∗ , F(x)) is the distance between F∗ and F(x), which may be deﬁned as F n ∑ |Fi∗ − Fi (x)| (Manhattan distance) i=1 v u nF ∗ D(F , F(x)) = u∑ (A.110) t [F ∗ − F (x)]2 (Euclidean distance) i i i=1 max |F ∗ − Fi (x)| (Chebychev distance) i

i

350

Optimization of Finite Dimensional Structures

A Pareto solution that is nearest to F∗ is found as the solution of Problem (A.109). The ideal point can be given using the independently minimized objective values, as illustrated in Fig. A.11(d), or it may be deﬁned simply by the preference of the decision maker. However, it is very diﬃcult to assign the ideal point appropriately. Therefore, the ideal point is modiﬁed interactively to search for the most preferred solution among the Pareto optimal solutions. To assist the decision maker in modiﬁcation of the ideal point, the trade-oﬀ ratio, e.g., ∆F2 /∆F1 in Fig. A.10(b) for nF = 2, can be found analytically utilizing the sensitivity coeﬃcients of the optimal solutions, which is called trade-oﬀ analysis. The method based on automatic trade-oﬀ analysis is called the aspiration level approach (Nakayama 1995).

A.5

Parametric structural optimization problem

A structural optimization problem sometimes depends on a problem parameter that is diﬀerent from the design variables. For example, the optimal solution of a sizing problem depends on the nodal locations. Therefore, it is important to investigate the sensitivity of the optimal solution with respect to parameters. In this section, we summarize approach called parametric programming (Frank 1978; Fiacco 1983; Gal and Greenberg 1997), which was originally developed for postoptimal analysis of linear programming problems. Application of the parametric programming approach is demonstrated for multidisciplinary optimization in Sec. 5.3 and second-order postoptimal analysis of seismic design in Sec. 6.7. The objective function to be minimized is a function of the design variable vector x, which is denoted by the superscript ( · )0 as F 0 (x). The equality and inequality constraints are given as G0i (x, u(x)) = 0 (i = 1, . . . , nE ) and Hi0 (x, u(x)) ≤ 0 (i = 1, . . . , nI ), respectively, where u(x) is the vector of nodal displacements, which are conceived as state variables. The optimization problem is formulated as Minimize

F 0 (x)

subject to

G0i (x, u(x)) = 0, Hi0 (x, u(x)) ≤ 0,

(A.111a) E

(A.111b)

I

(A.111c)

(i = 1, . . . , n ) (i = 1, . . . , n )

Note that the side constraints for x are supposed to be included in the general inequality constraints (A.111c). Let R0 (x, u(x)) denote the vector of constraint functions consisting of the equality constraints (A.111b) and the active inequality constraints in (A.111c) at an optimal solution. The ith component of R0 is denoted by Ri0 . The

Appendix

351

function of x only is deﬁned with a tilde as ei0 (x) = Ri0 (x, u(x)) R

(A.112)

The Karush-Kuhn-Tucker conditions for Problem (A.111) are written as n e0 ∂F 0 ∑ ∂ R j + λj = 0, (i = 1, . . . , m) ∂xi ∂x i j=1 A

(A.113)

where m is the number of variables, nA is the number of active constraints including the equality constraints, and λj (≥ 0) is the Lagrange multiplier. If e0 is related to the active inequality constraint, then λj ≥ 0 holds, whereas R j no restriction in sign exists for λj related to an equality constraint. e0 with respect to xi is obtained as The partial diﬀerentiation of R j n e0 ∂R ∂Rj0 ∑ ∂Rj0 ∂uk j = + , (i = 1, . . . , m; j = 1, . . . , nA ) ∂xi ∂xi ∂uk ∂xi

(A.114)

k=1

where n is the number of components in u. Let p = (p1 , . . . , pnP )⊤ denote a vector of nP parameters. The parametric form of Problem (A.111) is stated as Minimize

F (x, p)

(A.115a) E

subject to Gi (x, p) = 0, (i = 1, . . . , n )

(A.115b)

Hi (x, p) ≤ 0, (i = 1, . . . , n )

(A.115c)

I

where u(x) is not included in the arguments, for simplicity, because it can be regarded as an implicit function of x. The design variables, state variables, and the objective value at the optimal solution are functions of p, because they can be obtained for each speciﬁed value of p. The vector of active constraints R(x, p) is also redeﬁned as b R(p) = R(b x(p), p)

(A.116)

In the following, a function of p only is denoted by a hat as (c · ). The derivab tives of the optimal objective value F (p) = F (b x(p), p) with respect to pk are obtained from m ∑ ∂ Fb ∂F ∂F ∂b xi = + (A.117) ∂pk ∂pk i=1 ∂xi ∂pk bj (p) = 0 at an optimal solution remain Suppose that the active constraints R active for a small variation of pk ; i.e., bj ∂R = 0, (j = 1, . . . , nA ) ∂pk

(A.118)

352

Optimization of Finite Dimensional Structures

Then the following relation holds: ∂Rj ∑ ∂Rj ∂b xi + = 0, (j = 1, . . . , nA ) ∂pk ∂x ∂p i k i=1 m

(A.119)

By multiplying ∂b xi /∂pk on both sides of (A.113) and taking summation over i, we have m n m ∑ xi ∑ ∑ ∂Rj ∂b xi ∂F ∂b + =0 λj ∂xi ∂pk i=1 j=1 ∂xi ∂pk i=1 A

(A.120)

By multiplying λj on both sides of (A.119) and taking summation over j corresponding to the equality constraints and the active inequality constraints, we derive the following relation: n ∑ A

j=1

∂Rj ∑ ∑ ∂Rj ∂b xi + =0 λj ∂pk ∂x ∂p i k j=1 i=1 nA

λj

m

(A.121)

From (A.117), (A.120), and (A.121), ∑ ∂Rj ∂ Fb ∂F = + λj ∂pk ∂pk j=1 ∂pk nA

(A.122)

is derived. We can see from (A.122) that the derivative of the optimal objective value with respect to pk , which is called the parametric sensitivity coeﬃcient, can be obtained without computing the derivatives of the design variables and the Lagrange multipliers. Next, we compute the second-order derivatives of the optimal objective value Fb with respect to the parameters. Diﬀerentiation of (A.113) with respect to pk leads to m ∑ l=1

nA 2 2 ∑ ∂ R xl ∂ F ∂2F j ∂b + λj + ∂xi ∂xl j=1 ∂xi ∂xl ∂pk ∂xi ∂pk

n ∑ A

+

j=1

(

bj ∂Rj ∂ λ ∂ 2 Rj + λj ∂xi ∂pk ∂xi ∂pk

)

(A.123)

=0

bj and x The derivatives of λ bi are computed from a set of m+nA linear equations (A.120) and (A.123), where the ﬁrst term on the left-hand side of (A.123) consists of the Hessian of the Lagrangian.

Appendix

353

By further diﬀerentiating (A.122) with respect to pr , we obtain n m m ∑ ej ∂b ∂ 2 Fb ∂2F ∂ 2 F ∂b xi ∑ ∑ ∂2R xi + = + λj ∂pk ∂pr ∂pk ∂pr i=1 ∂pk ∂xi ∂pr j=1 i=1 ∂pk ∂xi ∂pr ( ) nA ∑ bj ∂ 2 Rj ∂Rj ∂ λ + λj + ∂pk ∂pr ∂pk ∂pr j=1 A

(A.124)

It is observed from (A.124) that the second-order derivative of Fb can be found bj and x without computing the second-order derivatives of λ bi , although their ﬁrst-order derivatives are needed.

A.6 A.6.1

Parametric curves and surfaces B´ ezier curve

As is seen in Secs. 6.4, 6.5, and 6.6, parametric curves and surfaces such as a B´ezier curve and a B´ezier surface can be eﬀectively used for modeling relatively complex curves and surfaces with a small number of variables. Parametric curves are also very eﬀective for generating a smooth optimal shape of a ¨ two-dimensional continuum (Braibant and Fleury 1984; Ozakca, Hinton, and Rao 1993; Eschenauer, Kobelev, and Schumacher 1994). In this section, we summarize the formulations of B´ezier curves and surfaces (Farin 1988; Farin, Hoschek, and Kim 2002). Note that other parametric representations, e.g., B-spline and non-uniform rational B-spline (NURBS) curves/surfaces, are not explained here, because they are not used in this book (see Farin, Hoschek, and Kim (2002) for more details). The B´ezier curve is deﬁned by the function called the Bernstein basis function or the Bernstein polynomial and the control points, as shown in Fig. A.12. The polygon that consists of the control points is called the control polygon or the deﬁning polygon. In Fig. A.12, the control polygon consists of the control points R0 , R1 , R2 , and R3 . The Bernstein basis function of order n is given as ( ) n i n Bi (t) = t (1 − t)n−i , (i = 0, 1, . . . , n) (A.125a) i ( ) n! for 0 ≤ i ≤ n n = (A.125b) i!(n − i)! i 0 for i < 0 or i > n where 0 ≤ t ≤ 1 is the parameter, and 00 = 0! = 1. For example, Bin (t) for

Optimization of Finite Dimensional Structures

354

R1

R2

R0

R3

FIGURE A.12: Control points and control polygon of a B´ezier curve of order 3. n = 3 are given as B03 (t) = (1 − t)3 , B13 (t) = 3t(1 − t)2 , B23 (t) = 3t2 (1 − t), B33 (t) = t3

(A.126)

which are plotted in Fig. A.13. Bin (t) can also be deﬁned by the following recursive form: n−1 Bin (t) = (1 − t)Bin−1 (t) + tBi−1 (t)

(A.127a)

B00 Bjn

=1

(A.127b)

= 0, for j < 0 or j > n

(A.127c)

For example, the basis functions of order 4 are obtained from Bi3 (t) as B04 (t) = (1 − t)B03 = (1 − t)4 , B14 (t) = (1 − t)B13 + tB03 = 3t(1 − t)3 + t(1 − t)3 = 4t(1 − t)3 , B24 (t) = (1 − t)B23 + tB13 = 3t2 (1 − t)2 + t2 (1 − t)2 = 6t2 (1 − t)2 , B34 (t) B44 (t)

= =

(1 − t)B33 tB33 = t4

+

tB23

(A.128)

= t (1 − t) + 3t (1 − t) = 4t (1 − t), 3

3

3

By using Bin (t), the B´ezier curve Pnb (t) of order n is given as Pnb (t)

=

n ∑

Ri Bin (t)

(A.129)

i=0

where Ri is the location vector of the ith control point. This way, the B´ezier curve of order n is deﬁned with n + 1 control points, and the order of the basis function is n. Examples of B´ezier curves are illustrated in Fig. A.14 for n = 3 with diﬀerent locations of the control point R3 . As is seen, the shape of the curve can be controlled by moving the control points. Note that these curves exist inside of the convex hull R0 R1 R3 R2 , indicated by dotted lines in Fig. A.14(a), formed by the control points.

Appendix

355

1 0.8 B 33

B 03 0.6

B 13

B 23

0.4 0.2 t

0 0

0.2

0.4

0.6

0.8

1

FIGURE A.13: Bernstein basis functions of order 3. R3

R1

R1

R3 R0

R0

R2

(a)

R2

(b)

FIGURE A.14: Relation between the control points and the shape of the B´ezier curve with R0 = (100, 100), R1 = (150, 200), R2 = (250, 100); (a) R3 = (300, 150), (b) R3 = (300, 250). It can be easily observed from (A.125) as well as Fig. A.13 that only B0n is 1 and the other terms are 0 at t = 0. On the other hand, Bnn = 1 and the other terms are 0 at t = 1. Therefore, from (A.129), the curve coincides with the points R0 and Rn at t = 0 and 1, respectively. It can also be derived by diﬀerentiating (A.125a) with respect to t that the tangent vectors at both ends with t = 0 and 1, respectively, coincide with the direction of the control polygon, as seen in Figs. A.12 and A.14. For example, for n = 3, the derivatives of Bi3 (t) with respect to t, denoted by B˙ i3 (t), are obtained as B˙ 03 (t) = −3(1 − t)2 , B˙ 13 (t) = 3(1 − 3t)(1 − t), B˙ 23 (t) = 3t(2 − 3t), B˙ 33 (t) = 3t2

(A.130)

Therefore, we have B˙ 03 (0) = −3, B˙ 13 (0) = 3, B˙ 13 (0) = 0, B˙ 23 (0) = 0 B˙ 03 (1) = 0, B˙ 13 (1) = 0, B˙ 13 (1) = −3, B˙ 23 (1) = 3

(A.131)

356

Optimization of Finite Dimensional Structures

FIGURE A.15: A tensor product B´ezier surface of order 3 × 3. ˙ 3 (t) at t = 0 and 1: which leads to the tangent vector P b ˙ 3 (0) = 3(R1 − R0 ), P ˙ 3 (1) = 3(R3 − R2 ) P b b

(A.132)

Furthermore, the following equation is derived for any n from (A.125a) and (A.125b): n ∑ Bin (t) = 1 (A.133) i=0

which leads to the property that the curve exists in the convex hull of the control polygon. These properties are eﬀectively used for controlling the shape of the curve interactively by moving the control points.

A.6.2

B´ ezier surface

The surface deﬁned as the product of B´ezier curves is called the tensor product B´ezier surface. Let u and v denote the parameters in the two perpendicular directions. The tensor product B´ezier surface of order n × m is deﬁned as n ∑ m ∑ Pn,m (u, v) = Ri,j Bin (u)Bjm (v) (A.134) b i=0 j=0

where Ri,j (i = 0, . . . , n; j = 0, . . . , m) are the control points, and Bin (u) and Bjm (v) are the Bernstein basis functions of orders n and m, respectively, in the directions of u and v. A tensor product B´ezier surface of order 3 × 3 is illustrated in Fig. A.15, where the thick lines are the edges of the control polygon, which is also called the control net. Note that the numbers of control points as well as the orders of the Bernstein basis function in the u- and v-directions can be diﬀerent. As is easily seen from (A.134), the curve of ﬁxed parameter value u or v forms a B´ezier curve; therefore, the boundary of the surface is a B´ezier curve. Although the tensor product B´ezier surface is very simple, it may not be suitable for modeling a complex shape, including a surface with a circular

Appendix

357

FIGURE A.16: A triangular B´ezier patch of order 3. plan. In this case, the triangular B´ezier patch of order n, deﬁned as follows, can be eﬀectively used: ∑ n Pnt (u, v, w) = Rαβγ Bαβγ (A.135a) α+β+γ=n n Bαβγ

n! = uα v β w γ α!β!γ!

(A.135b)

where (u, v, w) are the barycentric coordinates 0 ≤ u, ≤ 1, 0 ≤ v, ≤ 1, 0 ≤ w, ≤ 1, u + v + w = 1, and the summation is made over α, β, γ = 0, . . . , n with α + β + γ = n. Alternatively, the variable w, which is dependent on u and v, is removed as Pnt (u, v) =

n n−α ∑ ∑

n Rαβ Bαβ

(A.136a)

α=0 β=0 n Bαβ =

n! uα v β (1 − u − v)n−α−β α!β!(n − α − β)!

(A.136b)

An example of a triangular B´ezier patch of order 3 is illustrated in Fig. A.16, where the thick lines show the control net, and the thin lines are the set of points with a constant value of u, v, or w = 1 − u − v.

A.6.3

Adjoint curve

Let P(t) denote a parametric curve, e.g., a B´ezier curve, which is regarded as a primary curve or progenitor curve (Barnhill 1994; Hoschek 1985; Hoschek and Wissel 1988). A curve deﬁned by parameter t, using the properties of P(t), is called an adjoint curve. For example, the hodograph deﬁned by the velocity vector of a particle in motion, as shown in Fig. A.17, is an adjoint curve (Katoh, Ohsaki, and Tani 2002). In this case, the adjoint curve Pa (t) to the progenitor curve P(t) is deﬁned as ˙ Pa (t) = αP(t)

(A.137)

Optimization of Finite Dimensional Structures

358

(a) Velocity vector

(b) Hodograph

FIGURE A.17: Velocity vector and hodograph.

TABLE A.1: Point 0 1 2 3 4 5

Control points of the B´ezier curves in Fig. A.18. Case 1 Case 2 Case 3 (100.0, 100.0) (100.0, 100.0) (210.0, 400.0) (150.0, 300.0) (220.0, 300.0) (290.0, 200.0) (200.0, 100.0) (200.0, 50.0) (10.0, 100.0) (250.0, 100.0) (250.0, 50.0) (490.0, 100.0) (300.0, 300.0) (230.0, 300.0) (210.0, 200.0) (350.0, 100.0) (350.0, 100.0) (290.0, 400.0)

where α is a speciﬁed scaling parameter. The oﬀset curve that exists at a constant distance from a curve is very important in various ﬁelds of engineering, e.g., numerical control machining. The oﬀset curve D(t) is also an adjoint curve, because it is deﬁned using the unit normal vector n0 (t) of P(t) as D(t) = P(t) + dn0 (t)

(A.138)

where d is the speciﬁed oﬀset value. The properties of the oﬀset curve can also be expressed in a parametric form; e.g., the tangent vector of D(t) is written as ˙ D(t) = P(t) + dn˙0 (t) (A.139) The oﬀset curves are extensively used for deﬁning the shapes of the arch-type trusses in Secs. 6.4 and 6.6. Note that the oﬀset curve is not a B´ezier curve even if P(t) is a B´ezier curve. A variety of algorithms have been proposed for generating an approximate parametric form of an oﬀset curve or surface. Furthermore, if d is too large or P(t) has a complex shape, a loop or intersection exists in the oﬀset curve. Examples of oﬀset curves of the progenitor curve deﬁned by the B´ezier curve of order 5 are shown in Fig. A.18. The locations of the control points for each case are listed in Table A.1, and the distance d between the progenitor and oﬀset curves is 30. As is seen, if the locations of the control points are not appropriate, there may exist a loop or intersection, as seen in Figs. A.18(b) and (c), respectively.

Appendix

(a) Case 1

359

(b) Case 2

(c) Case 3 FIGURE A.18: Oﬀset curves of B´ezier curves with various locations of control points; thick solid line: progenitor curve, thin solid line: oﬀset curve, dotted line: control polygon.

A.7 A.7.1

Response spectrum approach SRSS method

The responses of structures under seismic motions can be evaluated using a time-history analysis if the record of input acceleration is given. Although an accurate response can be computed by using time-history analysis for a single deterministic input motion, seismic motions are highly unpredictable in the design process of structures in civil engineering. Therefore, the design response spectrum is usually given for representing the statistical properties of the possible seismic motions at a speciﬁc construction site. In this section, we brieﬂy summarize the response spectrum approach for completeness of the book (see textbooks, e.g., Chopra (2001) for details). Let K and M denote the n × n stiﬀness matrix and mass matrix, respec-

360

Optimization of Finite Dimensional Structures

tively, where n is the number of degrees of freedom. The rth eigenvalue Ωr and eigenvector Φr are deﬁned by KΦr = Ωr MΦr , (r = 1, . . . , n)

(A.140)

where the eigenvectors are ortho-normalized with respect to the mass matrix as Φ⊤ (A.141) r MΦs = δrs , (r, s = 1, . . . , n) where δrs is the Kronecker delta. Let I denote the vector consisting of 1 for the components corresponding to the direction of the input seismic motion and 0 for the remaining components. Then the rth participation factor βr is given as βr = Φ⊤ r MI, (r = 1, . . . , n)

(A.142)

The rth damping ratio hr may be deﬁned, e.g., using stiﬀness-proportional damping with h∗1 speciﬁed for the ﬁrst mode; i.e., √ Ωr , (r = 1, . . . , n) hr = h∗1 (A.143) Ω1 In this case, the damping matrix C is deﬁned as C = α1 K with an appropriate coeﬃcient α1 . If Rayleigh damping is used, the damping ratios for two diﬀerent modes can be speciﬁed, and C is deﬁned as C = α1 K + α2 M with coeﬃcients α1 and α2 . The speciﬁed displacement response spectrum is denoted by SD (Ωr , hr ), which is a function of Ωr and hr . By using the square-root-of-sum-of-squares (SRSS) method, we can compute the mean value Ujmax of the maximum response of the displacement component Uj , as follows, under the set of seismic r motions that are compatible with SD (Ωr , hr ) denoted simply by SD : v u nD u∑ max r β Φ )2 Uj = t (SD (A.144) r r,j r=1

where Φr,j is the jth component of Φr , and the lowest nD modes are assumed to be used for evaluation of the responses. Note that the higher modes cannot be neglected, especially for long-span domes and structures with some symmetry conditions. Let εr,i denote the strain component of the ith member of a truss or frame corresponding to the mode Φr . Then the maximum response strain εmax of i the ith member is also evaluated using the SRSS method: v u nD u∑ max r β ε )2 (A.145) εi = t (SD r r,i r=1

Appendix

361

The pseudovelocity response spectrum SV (Ωr , hr ) and the pseudoacceleration response spectrum SA (Ωr , hr ) are deﬁned as √ r SV = SV (Ωr , hr ) = Ωr SD (Ωr , hr ), (r = 1, . . . , n) (A.146a) r SA = SA (Ωr , hr ) = Ωr SD (Ωr , hr ), (r = 1, . . . , n)

(A.146b)

Then the maximum nodal velocity vimax and acceleration amax corresponding i to the ith displacement component are evaluated as v u nD u∑ max r β Φ )2 , vi = t (SV r r,i r=1

amax i

v u nD u∑ r β Φ )2 = t (SA r r,i

(A.147)

r=1

Note that the pseudodisplacement response spectrum is to be deﬁned as r = SD

1 r S Ωr A

(A.148)

r if SA is speciﬁed as the design response spectrum.

A.7.2

CQC method

It is well known that the accuracy of the SRSS method deteriorates if the eigenvalues, or frequencies, of the dominant modes are closely spaced, because the SRSS method assumes that the maximum modal responses are not correlated. For this case, the maximum seismic responses may be evaluated using the complete quadratic combination (CQC) method (Wilson, Der Kiureghian, and Bayo 1982), which is an extension of the SRSS method. The CQC method incorporates the correlation between the responses of the dominant modes, where the coeﬃcient τrs between the rth and sth modes is deﬁned by √ 3/2 8 hr hs (hr + αrs hs )αrs τrs = (A.149) 2 )2 + 4h h α (1 + α2 ) + 4(h2 + h2 )α2 (1 − αrs r s rs rs r s rs √

where

Ωs , (r, s = 1, . . . , n) (A.150) Ωr Then, for example, the maximum response strain εmax of the ith member is i found from v u nD nD u∑ ∑ max [SD (Ωr , hr )βr εr,i ]τrs [SD (Ωs , hs )βs εs,i ] εi =t (A.151) αrs =

r=1 s=1

362

Optimization of Finite Dimensional Structures

As is seen from (A.147) and (A.151), the CQC method reduces to the SRSS method if αrs = δrs with δrs being the Kronecker delta. The CQC method may also be used if multiple components are considered for input motions. It is assumed here that the three orthogonal directions denoted by directions 1, 2, and 3 correspond to the principal axes of the (j) structure. The participation factor βr of the rth mode for input in the jth direction is deﬁned as (j) βr(j) = Φ⊤ (A.152) r MI where I(j) is a vector of which the components corresponding to the displacement in the jth direction are 1, and the remaining components are 0. (j) Let SD (Ωr , hr ) denote the speciﬁed displacement response spectrum for the excitation in the jth direction. Then, the maximum response strain εmax i of the ith member is found from the following equation (Semby and Der Kiureghian 1985): v u 3 nD nD u∑ ∑ ∑ (j) (j) (j) (j) max =t εi [SD (Ωr , hr )βr εr,i ]τrs [SD (Ωs , hs )βs εs,i ] (A.153) j=1 r=1 s=1

A.7.3

Design response spectrum

The response spectrum by Newmark and Hall (1982) is used in most of the examples of seismic optimization in this book. The maximum values for acceleration, velocity, and displacement of the ground motion are denoted by CA , CV , and CD , respectively. The design displacement response spectrum, which (1) (5) is a function of Ωr , is deﬁned as the minimum values among SD , . . . , SD that are deﬁned as (1)

SD (Ωr ) = CA /Ωr (2) SD (Ωr ) (3) SD (Ωr ) (4) SD (Ωr ) (5) SD (Ωr )

=

16.2CA AA Ω−1.36 r

(A.154a) (A.154b)

= CA AA /Ωr √ = CV AV / Ωr

(A.154c) (A.154d)

= CD AD

(A.154e)

where the ampliﬁcation factors AA , AV , and AD are given as AA = 3.21 − 0.68 log(100hr ) AV = 2.31 − 0.41 log(100hr )

(A.155a) (A.155b)

AD = 1.82 − 0.27 log(100hr )

(A.155c)

The pseudovelocity response spectrum for the standard parameter values CA = 2.01 m/s2 , CV = 0.25 m/s, and CD = 0.1875 m is plotted in Fig. A.19 with the damping ratio hr = 0.02.

Appendix

363

Response velocity (m)

0.6

S4

0.5

S5

0.4

S3

0.3 0.2 0.1 0.0

0.1

1.0

10.0

Natural period (sec) FIGURE A.19:

A.7.4

Pseudovelocity response spectrum by Newmark and Hall.

Sensitivity analysis of mean maximum response

Let x denote the design variable representing, e.g., the cross-sectional area of a member of a truss and a nodal coordinate of a frame. The sensitivity coeﬃcients of Ωr and Φr are obtained from (2.39) and (2.42) in Sec. 2.3. Using (A.141) and diﬀerentiating (A.142), we obtain the sensitivity coeﬃcient of βr as ( )⊤ ∂βr ∂M ∂Φr = Φ⊤ I + MI (A.156) r ∂x ∂x ∂x Suppose, for simplicity, the responses are evaluated by using the SRSS r r method. Denoting SD β Φr,j by Ujr , the following relation is derived by taking the squares of both sides of (A.144) and diﬀerentiating them with respect to x: ( r nD [ ∑ ∂Ujmax ∂SD max r Uj = Uj βr Φr,j ∂x ∂x r=1 (A.157) )] ∂Φ ∂β r,j r r r Φr,j + SD βr + SD ∂x ∂x from which the sensitivity coeﬃcient of Ujmax is obtained. The sensitivity r coeﬃcient of SD = SD (Ωr , hr ) is easily obtained as r r r ∂SD ∂Ωr ∂SD ∂hr ∂SD = + ∂x ∂Ωr ∂x ∂hr ∂x

(A.158)

Therefore, if hr is given as a function of Ωr , then the sensitivity coeﬃcient of the maximum response can be analytically obtained. For the case of stiﬀness-

364

Optimization of Finite Dimensional Structures

List of column sections; A (×104 I (×10 mm ), Z (×106 mm3 ), Z p (×106 mm3 ). size A I Z 1 − 500 × 12 2.268 8.84 3.540 2 − 500 × 16 2.966 11.30 4.510 3 − 500 × 19 3.470 13.00 5.180 4 − 500 × 22 3.957 14.50 5.800 5 − 500 × 25 4.428 15.90 6.360 6 − 500 × 28 4.883 17.20 6.870 7 − 500 × 32 5.463 18.70 7.470 8 − 500 × 36 6.014 20.00 7.990

TABLE A.2: 8

mm2 ),

4

Zp 4.100 5.290 6.130 6.920 7.660 8.360 9.210 9.970

proportional damping, the following relation is derived by taking the squares of both sides of (A.143) and diﬀerentiating them with respect to x: 2hr

A.8

∂hr ∂Ωr ∂Ωr Ω1 + h2r = (h∗1 )2 ∂x ∂x ∂x

(A.159)

List of available standard sections of beams and columns

The list of available standard sections of steel members can be found from speciﬁcations, e.g., AISC in US and JIS in Japan. The members are classiﬁed into the sets, or series, with the same height or width of the section. However, it is not convenient to use the complete list of members of a set or members in diﬀerent sets for optimization, because the second moment of inertia I, the section modulus Z, and the plastic modulus Z p are not always increasing functions of the cross-sectional area A. Therefore, in structural optimization, several selected sections should be used so that I, Z, and Z p are listed in increasing order of A. The lists in Tables A.2 and A.3 for columns and beams, respectively, are used in the examples in Sec. 5.4. The values of I, Z, and Z p can be approximated by continuous functions of A (Grierson and Lee 1984; Sadek 1992; Wang and Arora 2006a; Pan, Ohsaki, and Kinosh*ta 2007). The standard method of least squares may be used for this purpose. However, normalization of A and I, which have diﬀerent magnitudes and units, may lead to the convergence of the parameter optimization process, as described below. For example, I is approximated as I(A) =a I0

(

A A0

)b (A.160)

Appendix

List of beam sections; A (×104 mm2 ), I (×10 mm ), Z (×106 mm3 ), Z p (×106 mm3 ). size A I Z 1 H − 500 × 200 × 9 × 12 0.923 3.750 1.500 2 H − 500 × 200 × 9 × 16 1.076 4.600 1.840 3 H − 500 × 200 × 9 × 19 1.190 5.210 2.090 4 H − 500 × 200 × 9 × 22 1.305 5.810 2.330 5 H − 500 × 200 × 12 × 22 1.442 6.050 2.420 6 H − 500 × 250 × 9 × 22 1.525 7.070 2.830 7 H − 500 × 250 × 12 × 22 1.662 7.310 2.920 8 H − 500 × 250 × 12 × 25 1.804 8.040 3.220 9 H − 500 × 250 × 12 × 25 1.947 8.750 3.500

365

TABLE A.3: 8

4

Zp 1.720 2.080 2.340 2.600 2.760 3.010 3.290 3.610 3.930

TABLE A.4: Parameters and approximation errors for a column and a beam. Column a b R I 1.030 0.819 2.545 × 10−3 Z 1.028 0.8189 2.474 × 10−3 p Z 1.016 0.9017 8.698 × 10−4 Beam a b R I 1.038 1.090 1.762 × 10−2 Z 1.039 1.088 1.840 × 10−2 Zp 1.019 1.077 3.832 × 10−3

where A0 and I0 are the speciﬁed values for normalization. Then the coeﬃcients a and b are obtained by minimizing the error R deﬁned as [ ( )b ]2 nL ∑ A I(A) R= −a I0 A0 i=1

(A.161)

where nL is the number of sections in the list, and Z and Z p are approximated similarly. Accuracy may be improved if we add the constant term to the righthand side of (A.160); however, optimization often does not converge because the solution turns out to be non-unique. Optimization library SNOPT Ver. 7.2 (Gill, Murray, and Saunders 2002) was used for ﬁnding the optimal parameter values. The smallest values in the list are used for A0 , I0 , Z0 , and Z0p for normalization. The minimum errors and the corresponding parameter values for a beam and a column in Tables A.2 and A.3 are listed in Table A.4. The approximate relation for the beam sections is plotted in Fig. A.20, which shows almost linear approximation. Alternatively, the cross-sectional properties can be approximated as linear functions of the reciprocal of Ai (Grierson and Chan 1993). It is also possible

Optimization of Finite Dimensional Structures

366

Second moment of inertia

2 1.8 1.6 1.4 1.2 1 1

1.2

1.4

1.6

1.8

2

Cross-sectional area

FIGURE A.20: Approximation of second moment of inertia of a beam.

to interpolate Ii and Zi as piecewise continuously diﬀerentiable functions of Ai (Sedaghati and Esmailzadeh 2003), as demonstrated in Example 5.1 in Sec. 5.1.

References

Aarts, E. and J. Korst (1989). Simulated Annealing and Boltzmann Machines: A Stochastic Approach to Combinatorial Optimization and Neural Computing. Chichester, UK: Wiley. Aarts, M. and J. K. Lenstra (Eds.) (1997). Local Search in Combinatorial Optimization. Chichester, UK: Wiley. Abrahamson, N. A., J. F. Schneider, and J. Stepp (1991). Empirical spatial coherency functions for application to soil-structure interaction analysis. Earthquake Spectra 21, 1–28. Achtziger, W. (1996). Truss topology optimization including bar properties diﬀerent for tension and compression. Struct. Opt. 12, 63–74. Achtziger, W. (1997). Topology optimization of discrete structures. In Topology Optimization in Structural Mechanics, CISM Courses and Lectures 374, pp. 57–100. Vienna: Springer-Verlag. Achtziger, W. (1999a). Local stability of trusses in the context of topology optimization, Part I: Exact modelling. Struct. Opt. 17, 235–246. Achtziger, W. (1999b). Local stability of trusses in the context of topology optimization, Part II: A numerical approach. Struct. Opt. 17, 247–258. Achtziger, W. (2007). On simultaneous optimization of truss geometry and topology. Struct. Multidisc. Optim. 33, 285–304. Achtziger, W., M. P. Bendsøe, A. Ben-Tal, and J. Zowe (1992). Equivalent displacement based formulations for maximum strength truss topology design. Impact of Computing in Science and Engineering 4 (4), 315–345. Achtziger, W. and M. Koˇcvara (2007a). On the maximization of the fundamental eigenvalue in topology optimization. Struct. Multidisc. Optim. 34, 181–195. Achtziger, W. and M. Koˇcvara (2007b). Structural topology optimization with eigenvalues. SIAM J. Optim. 18 (4), 1129–1164. Achtziger, W. and M. Stolpe (2008). Global optimization of truss topology with discrete bar areas, Part I: Theory of relaxed problems. Comput. Optim. Appl. 40 (2), 247–280. Achtziger, W. and M. Stolpe (2009). Global optimization of truss topology with discrete bar areas, Part II: Implementation and numerical results. Comput. Optim. Appl. 44 (2), 315–341.

367

368

Optimization of Finite Dimensional Structures

Adeli, H. (Ed.) (1994). Advances in Design Optimization. London: Chapman & Hall. Adeli, H. and H.-H. Chyou (1987). Microcomputer-aided optimal plastic design of frames. J. Computing in Civil Eng. 1 (1), 20–34. Adelman, H. M. and R. T. Haftka (1986). Sensitivity analysis of discrete structural systems. AIAA J. 24 (5), 823–832. Agrawal, R., T. Imielinski, and A. Swami (1998). Mining association rules between sets of items in large databases. In Proc. ACM SIGMOD Conf. on Management of Data (SIGMOD ’98), New York, pp. 207–216. ACM Press. AISC (1999). Load and Resistance Factor Design Speciﬁcation for Structural Steel Buildings. Chicago: American Institute of Steel Construction. Akg¨ un, M. A., J. H. Garcelon, and R. T. Haftka (2001). Fast exact linear and non-linear structural reanalysis and the Sherman–Morrison– Woodbury formulas. Int. J. Numer. Methods Eng. 50, 1587–1606. Allaire, G., F. Jouve, and H. Maliot (2004). Topology optimization for minimum stress design with the hom*ogenization method. Struct. Multidisc. Optim. 28, 87–98. Ant´ onio, C. A. C. (2002). A multilevel genetic algorithm for optimization of geometrically nonlinear stiﬀened composite structures. Struct. Multidisc. Optim. 24, 372–386. Argyris, J. H. and R. Roy (1972). General treatment of structural modiﬁcation. J. Struct. Div. ASCE 98 (ST2), 465–492. Armentano, V. A. and J. C. Arroyo (2004). An application of a multiobjective tabu search algorithm to a bicriteria ﬂowshop problem. J. Heuristics 10, 463–481. Arora, J. S. (2002). Methods for discrete variable structural optimization. In S. Burns (Ed.), Recent Advancements in Optimal Structural Design, pp. 1–40. Raton, VA: ASCE Press. Arora, J. S. (2004). Introduction to Optimum Design (2nd ed.). New York: Academic Press. Arora, J. S. (Ed.) (2007). Optimization of Structural and Mechanical Systems. Singapore: World Scientiﬁc. Arora, J. S., E. J. Haug, and K. Rim (1975). Optimal design of plane frames. J. Struct. Div. ASCE 101 (ST10), 2063–2078. Arora, J. S., M. W. Huang, and C. C. Hsieh (1994). Methods for optimization of nonlinear problems with discrete variables. Struct. Opt. 8 (2–3), 69–85. Arora, J. S. and C. H. Tseng (1987). IDESIGN User’s Manual, Ver. 3.5. ODL-87.1, Technical Report, Optimal Design Laboratory, The University of Iowa.

References

369

Arora, J. S. and Q. Wang (2005). Review of formulations for structural and mechanical system optimization. Struct. Multidisc. Optim. 30, 251–272. Atkinson, A. E. (1989). An Introduction to Numerical Analysis (2nd ed.). New York: John Wiley & Sons. Avil´es, R., M. B. G. Ajuria, J. Vallejo, and A. Hern´andez (1997). A procedure for the optimal synthesis of planar mechanisms based on non-linear position problems. Int. J. Numer. Methods Eng. 40, 1505–1524. Avis, D., N. Katoh, M. Ohsaki, I. Streinu, and S. Tanigawa (2007). Enumerating non-crossing minimally rigid frameworks. Graphs and Combinatorics 23(Suppl), 117–134. Balling, R. J. (1991). Optimal steel frame design by simulated annealing. J. Struct. Eng. 117 (6), 1780–1795. Balling, R. J., K. S. Pister, and V. Ciampi (1983). Optimal seismic-resistant design of a planar steel frame. Earthquake Eng. Struct. Dyn. 11, 541– 556. Balling, R. J. and J. Sobieszczanski-Sobieski (1996). Optimization of coupled systems: A critical overview of approaches. AIAA J. 34 (1), 6–17. Banachiewicz, T. (1937). Zur Berechungung der Determinanten, wie auch der Inversen, und zur darauf basierten Auﬂ¨osung der Systeme linearer Gleichungen. Acta Astronomica, S´er. C 3, 41–67. Barnhill, R. E. (Ed.) (1994). Geometry Processing for Design and Manufacturing. Philadelphia, PA: SIAM. Barthelemy, J. and J. Sobieszczanski-Sobieski (1983). Optimum sensitivity derivatives of objective functions in nonlinear programming. AIAA J. 21, 913–915. Barthold, F.-J. and N. Gerzen (2009). Application of singular value decomposition in structural optimization. In Proc. 8th World Congress of Structural and Multidisciplinary Optimization (WCSMO8), Paper No. 1392, Lisbon. Baˇzant, Z. P. and L. Cedolin (1991). Stability of Structures. New York: Oxford University Press. Baykasoglu, A. (2006). Applying multiple objective tabu search to continuous optimization problems with a simple neighborhood strategy. Int. J. Numer. Methods Eng. 65 (3), 406–424. Baykasoglu, A., S. Owen, and N. Gindy (1999a). Solution of goal programming models using a basic taboo search. J. Operational Research Society 50 (9), 960–973. Baykasoglu, A., S. Owen, and N. Gindy (1999b). A taboo search based approach to ﬁnd the Pareto optimal set in multiple objective optimization. Eng. Opt. 31, 731–748.

370

Optimization of Finite Dimensional Structures

Beckers, M. and C. Fleury (1997). A primal-dual approach in truss topology optimization. Comput. Struct. 64, 77–88. Ben-Israel, A., A. Ben-Tal, and S. Zolbec (1981). Optimality in Nonlinear Programming: A Feasible Direction Approach. New York: John Wiley & Sons. Ben-Tal, A. and M. P. Bendsøe (1993). A new method for optimal truss topology design. SIAM J. Optim. 3 (2), 322–358. Ben-Tal, A., F. Jarre, M. Koˇcvara, A. Nemirovski, and J. Zowe (2000). Optimal design of trusses under a nonconvex global buckling constraint. Opt. Eng. 1, 189–213. Ben-Tal, A., M. Koˇcvara, and J. Zowe (1993). Two nonsmooth approaches to simultaneous geometry and topology design of trusses. In M. P. Bendsøe and C. M. Mota Soares (Eds.), Topology Optimization of Structures, pp. 31–42. Dordrecht, The Netherlands: Kluwer Academic Publishers. Ben-tal, A. and A. Nemirovski (1994). Potential reduction polynomial time method for truss topology design. SIAM J. Optim. 4 (3), 596–612. Ben-Tal, A. and A. Nemirovski (1997). Robust truss topology optimization via semideﬁnite programming. SIAM J. Optim. 7, 991–1016. Bendsøe, M. P. (1989). Optimal shape design as a material distribution problem. Struct. Opt. 1, 193–202. Bendsøe, M. P., A. Ben-Tal, and J. Zowe (1994). Optimization method for truss geometry and topology design. Struct. Opt. 7, 141–159. Bendsøe, M. P. and O. Sigmund (2003). Topology Optimization: Theory, Methods and Applications. Berlin: Springer. Bennage, W. A. and A. K. Dhingra (1995). Optimization of truss topology using tabu search. Int. J. Numer. Methods Eng. 38, 4035–4052. Bennett, J. A. and M. E. Botkin (Eds.) (1986). The Optimum Shape: Automated Structural Design. New York: Plenum Press. Berke, L. and V. B. Venkayya (1974). Review of optimality criteria approaches to structural optimization. In J. L. A. Schmit (Ed.), Proc. Structural Optimization Symp., ASME AMD, Volume 1, pp. 23–34. Berry, M. J. A. and G. Linoﬀ (1997). Data Mining Techniques: For Marketing, Sales, and Customer Support. New York: John Wiley & Sons. Bersekas, D. P. (1982). Constrained Optimization and Lagrange Multiplier Methods. New York: Academic Press. Bhatti, M. A. and K. S. Pister (1981). A dual criteria approach for optimal design of earthquake resistant structural systems. Earthquake Eng. Struct. Dyn. 9, 557–572.

References

371

Bloebaum, C. L., P. Hajela, and J. Sobieszczanski-Sobieski (1992). Non-hierarchic system decomposition in structural optimization. Eng. Opt. 19, 171–186. Bochenek, B. and A. Gajewski (1986). Multimodal optimal design of a circular funicular arch with respect to in-plane and out-of-plane buckling. J. Struct. Mech. 14, 257–274. Bojczuk, D. and Z. Mr´oz (1998a). On optimal design of supports in beam and frame structures. Struct. Opt. 16, 47–57. Bojczuk, D. and Z. Mr´oz (1998b). Optimal design of trusses with account for topology variation. Mech. Struct. & Mach. 26 (1), 21–40. Bojczuk, D. and Z. Mr´oz (1999). Optimal topology and conﬁguration design of trusses with stress and buckling constraints. Struct. Multidisc. Optim. 17, 25–35. Box, M. J. (1965). A new method of constrained optimization and comparison with other methods. Computing J. 8, 42–52. Boyd, S. and L. Vandenberghe (2004). Convex Optimization. Cambridge, UK: Cambridge University Press. Bradt, M. (Ed.) (1986). Criteria and Methods of Structural Optimization. Dordrecht, The Netherlands: Martinus Nijhoﬀ Publishers. (ﬁrst edition in Polish, 1984). Braibant, V. and C. Fleury (1984). Shape optimal design using B-splines. Comput. Methods Appl. Mech. Engrg. 44, 247–267. Brigham, J. C. and W. Aquino (2007). Surrogate-model accelerated random search algorithm for global optimization with application to inverse material identiﬁcation. Comput. Methods Appl. Mech. Engrg. 196, 4561– 4576. Bruggi, M. (2008). On an alternative approach to stress constraints relaxation in topology optimization. Struct. Multidisc. Optim. 36, 125–141. Buhl, T., C. B. W. Pedersen, and O. Sigmund (2000). Stiﬀness design of geometrically nonlinear structures using topology optimization. Struct. Multidisc. Optim. 19, 93–104. Burns, S. (Ed.) (2002). Recent Advancements in Optimal Structural Design. Raton, VA: ASCE Press. Calladine, C. R. and S. Pellegrino (1991). First-order inﬁnitesimal mechanisms. Int. J. Solids Struct. 27 (4), 505–515. Canyurt, O. E. and P. Hajela (2005). A cellular framework for structural analysis and optimization. Comput. Methods Appl. Mech. Engrg. 194, 3516–3534. Cerny, V. (1985). Thermodynamical approach to the travelling salesman problem: An eﬃcient simulation algorithm. J. Optimization Theory Appl. 45, 41–51.

372

Optimization of Finite Dimensional Structures

Chan, C.-M., D. E. Grierson, and A. N. Sherbourne (1995). Automatic optimal design of tall steel building frameworks. J. Struct. Eng. 121 (5), 838–847. Chan, H. S. Y. (1969). On Foulkes mechanism in portal frame design for alternative loads. J. Appl. Mech. 36, 73–75. Chen, S. H., X. M. Wu, and Z. J. Yang (2006). Eigensolution reanalysis of modiﬁed structures using epsilon-algorithm. Int. J. Numer. Methods Eng. 66, 2115–2130. Chen, S. H., X. W. Yang, and H. D. Lian (2000). Comparison of several eigenvalue reanalysis methods for modiﬁed structures. Struct. Multidisc. Optim. 20, 253–259. Chen, T.-Y. and C.-J. Chen (1997). Improvements of simple genetic algorithms in structural design. Comput. Methods Appl. Mech. Engrg. 40, 1323–1334. Chen, W.-F. (1991). Stability Design of Steel Frames. Boca Raton, FL: CRC Press. Chen, W. F. and S. E. Kim (1997). LRFD Steel Design using Advanced Analysis. Boca Raton, FL: CRC Press. Cheng, F. Y. (2002). Multiobjective optimum design of seismic-resistant structures. In S. Burns (Ed.), Recent Advancements in Optimal Structural Design, pp. 241–255. Raton, VA: ASCE Press. Cheng, G. (1995). Some aspects of truss topology optimization. Struct. Opt. 10, 173–179. Cheng, G. and X. Guo (1997). ε-relaxed approach in structural topology optimization. Struct. Opt. 13, 258–266. Cheng, G. and Z. Jiang (1992). Study on topology optimization with stress constraints. Eng. Opt. 20, 129–148. Chern, J.-M. and W. Prager (1972). Optimal design of trusses for alternative loads. Ing.-Arch. 41, 225–231. Choi, K. K. and N.-H. Kim (2004). Structural Sensitivity Analysis and Optimization 1, Linear Systems. New York: Springer. Choi, K. K. and S.-L. Twu (1988). On equivalence of continuum and discrete methods of shape sensitivity analysis. AIAA J. 27 (10), 1418–1424. Chopra, A. K. (2001). Dynamics of Structures: Theory and Applications to Earthquake Engineering. Englewood Cliﬀs, NJ: Prentice-Hall. Cilly, F. H. (1900). The exact design of statically indeterminate frameworks, an exposition of its possibility but futility. Trans. ASCE 43, 353–407. Clough, R. W. and J. Penzien (1975). Dynamics of Structures. New York: McGraw-Hill.

References

373

Coello Coello, C. A. and G. T. Pulido (2005). Multiobjective structural optimization using a microgenetic algorithm. Struct. Multidisc. Optim. 30, 388–403. Coello Coello, C. A., G. T. Pulido, and M. S. Lechuga (2004). Handling multiple objectives with particle swarm optimization. IEEE Trans. Evolutionary Computation 8 (3), 256–279. Cohon, J. L. (1978). Multiobjective Programming and Planning. Mathematics in Science and Engineering, 140. New York: Academic Press. Cramer, E. J., J. E. Dennis, Jr., P. D. Frank, R. M. Lewis, and G. R. Shubin (1994). Problem formulation for multidisciplinary optimization. SIAM J. Optim. 4 (4), 754–776. Crisﬁeld, M. A. (1991). Non-Linear Finite Element Analysis of Solids and Structures, I: Essentials. New York: Wiley. Culmann, K. (1875). Die graphische Statik. Z¨ urich: Meyer & Zeller. Czyzak, P. and A. Jaszkiewicz (1998). Pareto simulated annealing: A metaheuristic technique for multi-objective combinatorial optimization. J. Multi-Criteria Decision Analysis 7, 34–47. Dantzig, G. B. and P. Wolfe (1960). Decomposition principle for linear programs. Operational Research 8 (1), 101–111. Dasgupta, D. and Z. Michalewicz (Eds.) (1997). Evolutionary Algorithms in Engineering Applications. Berlin: Springer. de Boer, H. and F. van Keulen (2000). Reﬁned semi-analytical design sensitivities. Int. J. Solids Struct. 37, 6961–6980. de Klerk, E., C. Roos, and T. Terlaky (1995). Semi-deﬁnite problems in truss topology optimization. Technical Report No. 95-128, Delft University of Technology, Delft, The Netherlands. DebChaudhury, A. and G. D. Gazis (1988). Response of MDOF systems to multiple support seismic excitation. J. Eng. Mech. 114 (4), 583–603. Dems, K. and W. Gatkowski (1995). Optimal design of a truss conﬁguration under multiloading conditions. Struct. Opt. 9, 262–265. Deng, L. and M. Ghosn (2001). Pseudoforce method for nonlinear analysis and reanalysis of structural systems. J. Struct. Eng. 127 (5), 570–578. Der Kiureghian, A. (1980). Structural response to stationary excitation. J. Eng. Mech. Div. ASCE 106 (6), 1195–1213. Der Kiureghian, A. (1996). A coherency model for spatially varying ground motions. Earthquake Eng. Struct. Dyn. 25, 99–111. Der Kiureghian, A. and A. Neuenhofer (1991). A response spectrum method for multiple-support seismic excitations. Technical Report No. UCB/EERC-91/08, Earthquake Engineering Research Center, University of California, Berkeley, CA.

374

Optimization of Finite Dimensional Structures

Der Kiureghian, A. and A. Neuenhofer (1992). Response spectrum method for multiple-support seismic excitation. Earthquake Eng. Struct. Dyn. 21, 713–740. Dey, A. and V. Gupta (1998). Response of multiply supported secondary systems to earthquakes in frequency domain. Earthquake Eng. Struct. Dyn. 27, 187–201. Dhingra, A. K. and W. A. Bennage (1995). Discrete and continuous variable structural optimization using tabu search. Eng. Opt. 24, 177–196. Ding, Y. and B. J. D. Esping (1991). An improved multilevel optimization approach. Comput. Struct. 38 (5–6), 557–567. Dobbs, M. W. and R. B. Nelson (1975). Application of optimality criteria to automated structural design. AIAA J. 14 (10), 1436–1443. Dobbs, W. and L. P. Felton (1969). Optimization of truss geometry. J. Struct. Div. ASCE 95 (ST10), 2105–2119. Doi, K., M. Yoshimura, S. Nishiwaki, and K. Izui (2009). Optimization method for lifecycle designs of products having diﬀerent lifecycle characteristics. In Proc. 8th World Congress of Structural and Multidisciplinary Optimization (WCSMO8), Paper No. 1071, Lisbon. Dorn, W., R. Gomory, and H. Greenberg (1964). Automatic design of optimal structures. J. de Mecanique 3, 25–52. Doˇ gan, E. and M. P. Saka (2009). Particle swarm design optimization of moment resisting steel frames with semi-rigid connections to LRFDAISC. In Proc. 8th World Congress of Structural and Multidisciplinary Optimization (WCSMO8), Paper No. 1109, Lisbon. Drucker, D. C. and R. T. Shield (1961). Bounds on minimum weight design. Quart. Appl. Math. 15, 269–281. Dubois, D. and P. Fortemps (1999). Computing improved optimal solutions to max–min ﬂexible constraint satisfaction problems. European J. Operational Research 118, 95–126. Duysinx, P. and M. P. Bendsøe (1998). Topology optimization of continuum structures with local stress constraints. Int. J. Numer. Methods Eng. 43, 1453–1478. Duysinx, P. and O. Sigmund (1998). New developments in handling stress constraints in optimal material distribution. In Proc. 7th Symp. on Multidisciplinary Analysis and Optimization, Paper AIAA-98-4906, pp. 1501–1509. Ebenau, C., J. Rottsch¨ afer, and G. Thierauf (2005). An advanced evolutionary strategy with an adaptive penalty function for mixed-discrete structural optimization. Advances in Engineering Software 36, 29–38. Ekeland, I. and R. T´emam (1999). Convex Analysis and Variational Problems. Philadelphia, PA: SIAM.

References

375

Elishakoﬀ, I., R. T. Haftka, and J. Fang (1994). Structural design under bounded uncertainty–optimization with anti-optimization. Comput. Struct. 53 (6), 1401–1405. Elishakoﬀ, I. and M. Ohsaki (2010). Optimization and Anti-Optimization of Structures. London: Imperial College Press. Erdal, F. and M. P. Saka (2009). Optimum design of castellated beams using harmony search method. In Proc. 8th World Congress of Structural and Multidisciplinary Optimization (WCSMO8), Paper No. 1102, Lisbon. Eriksson, A. (2008). Optimization in target movement simulations. Comput. Methods Appl. Mech. Engrg. 197, 4207–4215. Eschenauer, H. A., V. V. Kobelev, and A. Schumacher (1994). Bubble method for topology and shape optimization of structures. Struct. Opt. 8, 42–51. Evgrafov, A. (2005). On globally stable singular truss topologies. Struct. Multidisc. Optim. 29, 170–177. Evgrafov, A. (2006). Simultaneous optimization of topology and geometry of ﬂow networks. Struct. Multidisc. Optim. 32, 99–109. Evgrafov, A. and M. Patriksson (2003). Stable relaxations of stochastic stress-constrained weight minimization problems. Struct. Multidisc. Optim. 25, 189–198. Farin, G. (1988). Curves and Surfaces for Computer Aided Geometric Design. Boston: Academic Press. Farin, G., J. Hoschek, and M.-S. Kim (2002). Handbook of Computer Aided Geometric Design. Amsterdam: North-Holland. Faux, I. D. and M. J. Pratt (1979). Computational Geometry for Design and Manufacture. Chichester, UK: Ellis Horwood. Fiacco, A. V. (1983). Introduction to Sensitivity and Stability Analysis in Nonlinear Programming. New York: Academic Press. Fiacco, A. V. and G. P. M. Cormic (1968). Nonlinear Programming: Sequential Unconstrained Minimization Techniques. New York: John Wiley & Sons. Fleury, C. (1979). A uniﬁed approach to structural weight minimization. Comput. Methods Appl. Mech. Engrg. 20, 17–38. Fleury, C. (1980). An eﬃcient optimality criteria approach to the minimum weight design of elastic structures. Comput. Struct. 11, 163–173. Fleury, C. (1989a). CONLIN: An eﬃcient dual optimizer based on convex approximation. Struct. Opt. 1 (2), 81–89. Fleury, C. (1989b). Eﬃcient approximation concepts using second order information. Int. J. Numer. Methods Eng. 28, 2041–2058.

376

Optimization of Finite Dimensional Structures

Floudas, C. A. (1995). Nonlinear and Mixed-Integer Optimization. New York: Oxford University Press. Foley, C. M. (2002). Optimized performance-based design of buildings. In S. Burns (Ed.), Recent Advancements in Optimal Structural Design, pp. 169–240. Raton, VA: ASCE Press. Forsythe, G. E. and W. R. Wasow (1960). Finite-Diﬀerence Methods for Partial Diﬀerential Equations. New York: Wiley. Foulkes, J. (1954). Minimum weight design of structural frames. Proc. Royal Soc., London, Ser. A 82, 492–494. Fox, R. L. and M. P. Kapoor (1968). Rates of change of eigenvalues and eigenvectors. AIAA J. 6, 2426–2429. Frank, P. M. (1978). Introduction to System Sensitivity Theory. New York: Academic Press. Fredricson, H. (2005). Topology optimization of frame structures – joint penalty and material selection. Struct. Multidisc. Optim. 30, 193–200. Fredricson, H., T. Johansen, A. Klarbring, and J. Petersson (2003). Topology optimization of frame structures with ﬂexible joints. Struct. Multidisc. Optim. 25, 199–214. Freeman, S. A. (2004). Review of the development of the capacity spectrum method. ISEL J. Earthquake Technology 41 (1), 1–13. Friedman, Z. and M. B. Fuchs (1987). Multilevel optimal design of thinwalled continuous beams. Comput. Struct. 25, 405–414. Frye, M. J. and G. A. Morris (1975). Analysis of frames with ﬂexibly connected steel frames. Canadian J. Civil Eng. 2 (3), 280–291. Fuchs, M. B. (1993). N th-order stiﬀness sensitivities in structural analysis. Struct. Opt. 5, 207–212. Fujisawa, K., M. Kojima, and K. Nakata (1997). SDPA (semideﬁnite programming algorithm). In Proc. 2nd Workshop on High Performance Optimization Techniques, Rotterdam, The Netherlands. Fujita, S. and M. Ohsaki (2009). Shape optimization of shells considering strain energy and algebraic invariants of parametric surface. In Proc. 9th Asian Paciﬁc Conf. on Shell and Spatial Struct. (APCS2009), Paper No. P0012, Nagoya, Japan. f*ckuda, T., K. Mori, and M. Tsukiyama (1999). Parallel search for multimodal function optimization with diversity and learning of immune algorithm. In Artiﬁcial Immune Systems and Their Applications. Berlin: Springer. Gal, T. (1979). Postoptimal Analyses, Parametric Programming, and Related Topics. New York: Mc-Graw Hill.

References

377

Gal, T. and H. J. Greenberg (Eds.) (1997). Advances in Sensitivity Analysis and Parametric Programming. Norwell, MA: Kluwer Academic Publishers. Galileo Galilei (1638). Discorsi e Dimonstrazioni Matematiche, Interno, a Due Nuove Scienze. Leida. Ganzerli, S., C. P. Pantelides, and L. D. Reaveley (2000). Performancebased design using structural optimization. Earthquake Eng. Struct. Dyn. 29, 1677–1690. Garcea, G., G. Formica, and R. Casciaro (2005). A numerical analysis of inﬁnitesimal mechanisms. Int. J. Numer. Methods Eng. 62, 979–1012. Gavarini, C. and D. Veneziano (1972). Minimum weight limit design under uncertainty. Meccanica 7 (2), 98–104. Geem, Z. W., J.-H. Kim, and G. V. Loganathan (2001). A new heuristic optimization algorithm: Harmony search. Simulation 76 (2), 60–68. Ghasemi, M. R. and M. Farshchin (2009). Multi-objective weight and eigenperiod optimization of steel moment frames under seismic conditions, using ant colony method. In Proc. 8th World Congress of Structural and Multidisciplinary Optimization (WCSMO8), Paper No. 1362, Lisbon. Giger, M. and P. Ermanni (2006). Evolutionary truss topology optimization using a graph-based parameterization concept. Struct. Multidisc. Optim. 32, 313–326. Gill, P. E., W. Murray, and M. A. Saunders (2002). SNOPT: An SQP algorithm for large-scale constrained optimization. SIAM J. Optim. 12, 979–1006. Glover, F. (1975). Improved linear integer programming formulations of nonlinear integer problems. Management Science 22 (4), 455–460. Glover, F. (1989). Tabu search: Part I. ORSA J. Computing 1 (3), 190–206. Glover, F. and M. Laguna (1997). Tabu Search. Boston: Kluwer Academic Publishers. Goldberg, D. E. (1989). Genetic Algorithms in Search, Optimization, and Machine Learning. Reading, MA: Addison-Wesley. Goldberg, D. E. and J. Richerdson (1987). Genetic algorithms with sharing for multimodal function optimization. In Proc. 2nd Int. Conf. Genetic Algorithms on Genetic Algorithms and Their Application, pp. 41–49. Hillsdale, NJ: L. Erlbaum Associates Inc. Gondzio, J. (1995). HOPDM – a fast LP solver based on a primal-dual interior point method. European J. Operational Research 85, 221–225. Graver, J., B. Servatius, and H. Servatius (1993). Combinatorial Rigidity. Graduate Studies in Mathematics 2. Providence: American Mathematical Society.

378

Optimization of Finite Dimensional Structures

Greiner, D., J. M. Emperador, and G. Winter (2004). Single and multiobjective frame optimization by evolutionary algorithms and the autoadaptive rebirth operator. Comput. Methods Appl. Mech. Engrg. 193, 3711–3743. Grierson, D. E. and C.-M. Chan (1993). An optimality criteria design method for tall steel buildings. Advances in Engineering Software 16, 119–125. Grierson, D. E. and C. W. Chiu (1984). Optimal synthesis of frameworks under multilevel performance constraints. Comput. Struct. 18 (5), 889– 898. Grierson, D. E. and W. H. Lee (1984). Optimal synthesis of steel frameworks using standard sections. J. Struct. Mech. 12 (3), 335–370. Grierson, D. E. and W. H. Pak (1993). Optimal sizing, geometrical and topological design using a genetic algorithm. Struct. Opt. 6, 151–159. Grierson, D. E. and L. A. Schmit (1982). Synthesis under service and ultimate performance constraints. Comput. Struct. 15 (4), 405–417. Guillet, S., F. No¨el, and J. C. L´eon (1996). Structural shape optimization of parts bounded by free-form surfaces. Struct. Opt. 11, 159–169. Gunnlaugsson, G. A. and J. B. Martin (1973). Optimality conditions for fully stressed designs. SIAM J. Appl. Math. 25 (3), 474–482. Guo, X. and G. Cheng (2000). An extrapolation approach for the solution of singular optima. Struct. Multidisc. Optim. 19, 255–262. Guo, X., G. Cheng, and K. Yamazaki (2001). A new approach for the solution of singular optima in truss topology optimization with stress and local buckling constraints. Struct. Multidisc. Optim. 22, 364–372. Gupta, V. (1997). Acceleration transfer function of secondary systems. J. Eng. Mech. 123 (7), 678–675. Gurav, S. P., M. Langhaar, J. F. L. Goosen, and F. van Keulen (2005). Uncertainty-based design optimization of MEMS structures using combined cycle-based alternating anti-optimization. In Proc. 6th World Congress of Structural and Multidisciplinary Optimization (WCSMO6), Rio de Janeiro. Haber, R. B. and J. F. Abel (1982). Initial equilibrium solution methods for cable reinforced membranes, Part I: Formulations. Comput. Methods Appl. Mech. Engrg. 30, 263–284. Haftka, R. T. (1985). Simultaneous analysis and design. AIAA J. 23 (7), 1099–1103. Haftka, R. T. and H. M. Adelman (1989). Recent developments in structural sensitivity analysis. Struct. Opt. 1, 137–151.

References

379

Haftka, R. T., Z. G¨ urdal, and M. P. Kamat (1990). Elements of Structural Optimization. Dordrecht, The Netherlands: Kluwer Academic Publishers. Haftka, R. T. and M. P. Kamat (1989). Simultaneous nonlinear structural analysis and design. Comp. Mech. 4, 409–416. Haftka, R. T., J. A. Nachlas, L. T. Watson, T. Rizzo, and R. Desai (1987). Two-point constraint approximation in structural optimization. Comput. Methods Appl. Mech. Engrg. 60, 289–301. Haftka, R. T. and J. H. Starnes, Jr. (1976). Application of a quadratic extended interior penalty function for structural optimization. AIAA J. 14, 718–724. Haftka, R. T. and L. T. Watson (2005). Multidisciplinary design optimization with quasiseparable subsystems. Optim. Eng. 6, 9–20. Hagish*ta, T. and M. Ohsaki (2007). Optimization of placement of braces for a frame with semi-rigid joints by scatter search. In Proc. 7th World Congress of Structural and Multidisciplinary Optimization (WCSMO7), Seoul, pp. 692–701. Hagish*ta, T. and M. Ohsaki (2008a). Optimal placement of braces for steel frames with semi-rigid joints by scatter search. Comput. Struct. 86, 1983–1993. Hagish*ta, T. and M. Ohsaki (2008b). Topology mining for optimization of framed structures. J. Advanced Mechanical Design, Systems, and Manufacturing, JSME 2 (3), 417–428. Hagish*ta, T. and M. Ohsaki (2009). Topology optimization of trusses by growing ground structure approach. Struct. Multidisc. Optim. 37 (4), 377–393. Hajela, P. (1997). Stochastic search in discrete structural optimization simulated annealing, genetic algorithm and neural networks. In W. Gutkowski (Ed.), Discrete Structural Optimization, CISM Courses and Lectures 373. Vienna: Springer-Verlag. Hajela, P. and E. Lee (1995). Genetic algorithms in truss topological optimization. Int. J. Solids Struct. 32 (22), 3341–3357. Hajela, P. and C.-Y. Lin (1991). Optimal design of viscoelastically damped beam structures. Appl. Mech. Rev. 44, 96–102. Hajela, P. and C.-Y. Lin (1992). Genetic search strategies in multicriterion optimal design. Struct. Opt. 4, 99–107. Hand, D., H. Mannila, and P. Smyth (2001). Principles of Data Mining. Cambridge, MA: MIT Press. Hansen, M. P. (1997). Tabu search for multiobjective optimization: MOTS. In Proc. 13th Int. Conf. on MCDM, Cape Town, pp. 6–10.

380

Optimization of Finite Dimensional Structures

Hao, H. and X. N. Duan (1995). Seismic response of asymmetric structures to multiple ground motions. J. Struct. Eng. 121 (11), 1557–1564. Haque, M. I. (1996). Optimal frame design with discrete members using the complex method. Comput. Struct. 59 (5), 847–858. Harichandran, R. S., A. Hawwari, and B. N. Swedian (1996). Response of long-span bridges to spatially varying ground motion. J. Struct. Eng. 122 (5), 476–484. Hart, W. (1998). Sequential stopping rules for random optimization methods with applications to multistart local search. SIAM J. Optim. 9 (1), 270–290. Harville, D. (1976). Extension of the Gauss–Markov theorem to include the estimation of random eﬀects. Ann. Statist. 4 (2), 384–395. Hasan, R., L. Xu, and D. E. Grierson (2002). Push-over analysis for performance-based seismic design. Comput. Struct. 80, 2483–2493. Haug, E. J. and J. Cea (Eds.) (1981). Optimization of Distributed Parameter Structures 1, 2. Alphen aan den Rijn, The Netherlands: Noordhoﬀ. Haug, E. J. and K. K. Choi (1982). Systematic occurrence of repeated eigenvalues in structural optimization. J. Optimization Theory Appl. 38, 251–274. Haug, E. J., K. K. Choi, and V. Komkov (1986). Design Sensitivity Analysis of Structural Systems. New York: Academic Press. Hayalioglu, M. S. (2000). Optimum design of geometrically non-linear elastic-plastic steel frames via genetic algorithm. Comput. Struct. 77, 527–538. Hayalioglu, M. S. and S. O. Degertekin (2005). Minimum cost design of steel frames with semi-rigid connections and column bases via genetic optimization. Comput. Struct. 83, 1849–1863. Hegemier, G. A. and W. Prager (1969). On Michell trusses. Int. J. Mech. Sci. 11, 209–215. Hemp, W. S. (1973). Optimum Structures. Oxford, UK: Clarendon Press. Henderson, C. R., O. Kempthorne, S. R. Searle, and C. M. von Krosigk (1959). The estimation of environmental and genetic trends from records subject to culling. Biometrics 15 (2), 192–218. Henderson, H. V. and R. Searle (1981). On deriving the inverse of a sum of matrices. SIAM Review 23, 53–60. Heyman, J. (1959). On the absolute minimum weight design of framed structures. Quart. J. Mech. Appl. Math. 12 (3), 314–324. Ho, S. L., S. Yang, H. C. Wong, and G. Ni (2003). A simulated annealing algorithm for multiobjective optimizations of electromagnetic devices. IEEE Transactions on Magnetics 39 (3), 1285–1288.

References

381

Holland, J. H. (1975). Adaptation in Natural and Artiﬁcial Systems. Ann Arbor, MI: The University of Michigan Press. Horn, R. A. and C. R. Johnson (1990). Matrix Analysis (Reprint ed.). Cambridge, MA: Cambridge University Press. Horst, R., P. M. Pardalos, and N. V. Thoai (1995). Introduction to Global Optimization. Dordrecht, The Netherlands: Kluwer Academic Publishers. Horst, R. and H. Tuy (1985). Global Optimization. Berlin: Springer-Verlag. Hoschek, J. (1985). Oﬀset curves in the plane. Comput.-Aided Design 17, 77–82. Hoschek, J. and N. Wissel (1988). Optimal approximate conversion of spline curves and spline approximation of oﬀset curves. Comput.-Aided Design 20, 475–483. Hotelling, H. (1943). Some new methods in matrix calculation. Ann. Math. Statist. 14, 1–34. Hsieh, C. C. and J. S. Arora (1984). Design sensitivity analysis and optimization of dynamic response. Comput. Methods Appl. Mech. Engrg. 43, 195–219. Hu, N. (1992). Tabu search method with random moves for globally optimal design. Int. J. Numer. Methods Eng. 35, 1055–1070. Hu, T. C. and R. T. Shield (1961). Uniqueness in the optimum design of structures. J. Appl. Mech. 83, 1–4. ILOG (2007). ILOG CPLEX 10.2 User’s Manual. ILOG Inc. Imai, K. and L. A. Schmit (1982). Conﬁguration optimization of trusses. J. Struct. Div. ASCE 107 (ST5), 745–756. Izmailov, A. F. and M. V. Solodov (2008). Mathematical programs with vanishing constraints: Optimality conditions, sensitivity, and a relaxation method. J. Optimization Theory Appl. 142 (3), 69–99. J´ armai, K., J. Farkas, and Y. Kurobane (2006). Optimum design of a multistorey steel frame. Engineering Structures 28, 1038–1048. Jarre, F., M. Koˇcvara, and J. Zowe (1998). Optimal truss design by interiorpoint methods. SIAM J. Optim. 8 (4), 1084–1107. Jenkins, W. M. (1991). Towards structural optimization via the genetic algorithm. Comput. Struct. 40, 1321–1327. Jilla, C. D. and D. W. Miller (2001). Assessing the performance of a heuristic simulated annealing algorithm for the design of distributed satellite systems. Acta Astronautica 48 (5), 529–543. Jog, C. S. and R. B. Haber (1996). Stability of ﬁnite element models for distributed-parameter optimization and topology design. Comput. Methods Appl. Mech. Engrg. 130, 203–226.

382

Optimization of Finite Dimensional Structures

Jones, D. F., S. K. Mirrazavi, and M. Tamiz (2002). Multi-objective metaheuristics: An overview of the current state-of-the-art. European J. Operational Research 137, 1–9. Jutte, C. V. and S. Kota (2008). Design of nonlinear springs for prescribed load-displacement functions. J. Mech. Design 130 (8), Paper–081403. Kaelo, P. and M. M. Ali (2006). Some variants of the controlled random search algorithm for global optimization. J. Optimization Theory Appl. 130 (2), 253–264. Kameshki, E. S. and M. P. Saka (2001a). Genetic algorithm based optimum bracing design of non-sway tall plane frames. J. Const. Steel Res. 57, 1081–1097. Kameshki, E. S. and M. P. Saka (2001b). Optimum design of nonlinear steel frames with semi-rigid connections using genetic algorithm. Comput. Struct. 79, 1593–1604. Kameshki, E. S. and M. P. Saka (2003). Genetic algorithm based optimum design of nonlinear planar steel frames with various semi-rigid connections. J. Const. Steel Res. 59, 109–134. Kaneko, I. and C. D. Ha (1983). A decomposition procedure for large-scale optimum plastic design problems. Int. J. Numer. Methods Eng. 19, 873– 889. Kanno, Y. and M. Ohsaki (2007). Maximization of minimal eigenvalue of structures by using sequential semideﬁnite programming. In Proc. 7th World Congress of Structural and Multidisciplinary Optimization (WCSMO7), Seoul, pp. 1121–1130. Kanno, Y., M. Ohsaki, and N. Katoh (2001). Sequential semideﬁnite programming for optimization of framed structures under multimodal buckling constraints. Int. J. Structural Stability and Dynamics 1 (4), 585– 602. Kanno, Y., M. Ohsaki, and N. Katoh (2002). Symmetricity of the solution of semideﬁnite programming. Struct. Multidisc. Optim. 24, 225–232. Kanno, Y., M. Ohsaki, K. Murota, and N. Katoh (2001). Group symmetry in interior-point methods for semi-deﬁnite program. Optim. Eng. 2, 293– 320. Kargahi, M., J. C. Andersen, and M. M. Dessouky (2007). Structural weight optimization of frames using tabu search, Part I: optimization procedure. J. Struct. Eng. 132 (12), 1858–1868. Karmarkar, N. (1984). A new polynomial-time algorithm for linear programming. Combinatorica 4, 373–395. Katoh, N., M. Ohsaki, and A. Tani (2002). Introduction to Architectural Systems. Creators Library 3. Tokyo: Kyoritsu Shuppan (in Japanese).

References

383

Kaveh, A. (1986). Statical bases for an eﬃcient ﬂexibility analysis of planar trusses. J. Struct. Mech. 14 (4), 475–488. Kaveh, A. (1991). Graph and structures. Comput. Struct. 40 (4), 893–901. Kaveh, A. (2004). Structural Mechanics: Graph and Matrix Methods (3rd ed.). Somerset, UK: Research Studies Press. Kaveh, A., B. F. Azar, and S. Talatahari (2008). Ant colony optimization for design of space trusses. Int. J. Space Structures 23 (3), 167–181. Kavlie, D., H. Graham, and G. H. Powell (1971). Eﬃcient reanalysis of modiﬁed structures. J. Struct. Div. ASCE 97 (ST1), 377–392. Kawamoto, A. (2005). Path-generation of articulated mechanisms by shape and topology variations in non-linear truss presentation. Int. J. Numer. Methods Eng. 64, 1557–1574. Kawamoto, A., M. P. Bendsøe, and O. Sigmund (2004a). Articulated mechanism design with a degree of freedom constraint. Int. J. Numer. Methods Eng. 61, 1520–1545. Kawamoto, A., M. P. Bendsøe, and O. Sigmund (2004b). Planar articulated mechanism design by graph theoretical enumeration. Struct. Opt. 27, 295–299. Kawamura, H., H. Ohmori, and N. Koto (2002). Truss topology optimization by a modiﬁed genetic algorithm. Struct. Multidisc. Optim. 23, 467– 472. Kennedy, J. (1997). The particle swarm: Social adaptation of knowledge. In Proc. Int. Conf. Evolutionary Computation, Piscataway, NJ, pp. 303– 308. IEEE. Khot, N. S., L. Berke, and V. B. Venkayya (1978). Comparison of optimality criteria algorithms for minimum weight design of structures. AIAA J. 17 (2), 182–190. Khot, N. S. and M. P. Kamat (1985). Minimum weight design of truss structures with geometric nonlinear behavior. AIAA J. 23, 139–144. Kicher, T. P. (1966). Optimum design-minimum weight versus fully stressed. J. Struct. Div. ASCE 92 (ST6), 265–279. Kicinger, R., T. Arciszewski, and K. de Jong (2005). Evolutionary computation and structural design: A survey of the state-of-the-art. Comput. Struct. 83, 1943–1978. Kim, H., R. T. Haftka, W. H. Mason, L. T. Watson, and B. Grossman (2002). Probabilistic modeling of errors from structural optimization based on multiple starting points. Optim. Eng. 3, 415–430. Kim, H. M., N. F. Michelena, P. Y. Papalambos, and T. Jiang (2003). Target cascading in optimal structural design. J. Mech. Design 125, 474–480.

384

Optimization of Finite Dimensional Structures

Kim, M.-J., G.-W. Jang, and Y. Y. Kim (2008). Application of a ground beam-joint topology optimization method for multi-piece frame structure design. J. Mechanical Design 130 (8), Paper–081401. Kim, Y. Y., G. W. Jang, J. H. Park, J. S. Hyun, and S. J. Nam (2005). Conﬁguration design of rigid link mechanisms by an optimization method: A ﬁrst step. In Proc. IUTAM Symp. on Topological Design Optimization of Structures, Machines and Materials, Dordrecht, The Netherlands, pp. 251–260. Springer. Kimura, T. and H. Ohmori (2008). Computational morphogenesis of free form shells. J. Int. Assoc. Shells and Spatial Struct. 49 (3), 175–180. Kirkpatrick, S., C. D. Gelatt, Jr., and M. P. Vecchi (1983). Optimization by simulated annealing. Science 220 (4598), 671–680. Kirsch, U. (1972). Optimum design by partitioning into substructures. J. Struct. Div. ASCE 98 (ST1), 249–267. Kirsch, U. (1975). Multilevel approach to optimum structural design. J. Struct. Div. ASCE 101 (ST4), 957–974. Kirsch, U. (1989a). Optimal topologies of truss structures. Appl. Mech. Rev. 42, 223–239. Kirsch, U. (1989b). Optimal topologies of truss structures. Comput. Methods Appl. Mech. Engrg. 72, 15–28. Kirsch, U. (1990). On singular topologies in optimum structural design. Struct. Opt. 2, 133–142. Kirsch, U. (1993). Eﬃcient reanalysis for topological optimization. Struct. Opt. 6, 143–150. Kirsch, U. (1994). Eﬃcient sensitivity analysis for structural optimization. Comput. Methods Appl. Mech. Engrg. 117, 143–156. Kirsch, U. (1995). Improved stiﬀness-based ﬁrst-order approximations for structural optimization. AIAA J. 33 (1), 143–150. Kirsch, U. (1996). Integration of reduction and expansion process in layout optimization. Struct. Opt. 11, 13–18. Kirsch, U. (2000). Combined approximation: A general reanalysis approach for structural optimization. Struct. Multidisc. Optim. 20, 97–106. Kirsch, U. and S. Liu (1995). Exact structural reanalysis by a ﬁrst-order reduced basis approach. Struct. Opt. 10, 153–158. Kirsch, U. and G. I. N. Rozvany (1994). Alternative formulations of structural optimization. Struct. Opt. 7, 32–41. Kirsch, U. and M. F. Rubinstein (1972). Structural reanalysis by iteration. Comput. Struct. 2, 497–510. Kishi, N. and W.-F. Chen (1990). Moment-rotation relations of semi-rigid connections with angles. J. Struct. Eng. 116 (7), 1813–1834.

References

385

Klarbring, A., J. Petersson, B. Torstenfelt, and M. Karlsson (2003). Topology optimization of ﬂow networks. Comput. Methods Appl. Mech. Engrg. 192, 3909–3932. Kleiber, M. (1997). Parameter Sensitivity in Nonlinear Mechanics. Chichester, UK: John Wiley & Sons. Kocer, F. Y. and J. S. Arora (1999). Optimal design of H-frame transmission poles for earthquake loading. J. Struct. Eng. 125 (11), 1299–1309. Kogiso, N., L. T. Watson, Z. G¨ urdal, and R. T. Haftka (1994). Genetic algorithm with local improvement for composite laminate design. Struct. Opt. 7, 207–218. Kojima, M., S. Shindoh, and S. Hara (1997). Interior-point methods for the monotone semideﬁnite linear complementarity problems. SIAM J. Optim. 7, 86–125. Koll´ ar, L. (Ed.) (1999). Structural Stability in Engineering Practice. London: E & FN Spon. Koski, J. and R. Silvennoisen (1987). Norm methods and partial weighting in multicriterion optimization of structures. Int. J. Numer. Methods Eng. 24, 1101–1121. Koˇcvara, M. (1997). Topology optimization with displacement constraints: A bilevel programming approach. Struct. Opt. 14, 256–263. Koˇcvara, M. and J. V. Outrata (2006). Eﬀective reformulation of the truss topology design problem. Optim. Eng. 7, 201–219. Koˇcvara, M. and J. Zowe (1995). How to optimize mechanical structures simultaneously with respect to topology and geometry. In Proc. 1st World Congress of Structural and Multidisciplinary Optimization (WCSMO1), Goslar, Germany, pp. 135–140. Kravanja, S., Z. Kravanja, and B. S. Bedenik (1998). The MINLP optimization approach to structural optimization, Part I: General view on simultaneous topology and parameter optimization. Int. J. Numer. Methods Eng. 43, 263–292. Kreisselmeier, G. and R. Steinhauser (1983). Application of vector performance optimization to a robust control loop design for a ﬁghter aircraft. Int. J. Control 37, 251–284. Krishna, P. (1979). Cable Suspended Roofs. New York: Mc-Graw Hill. Krishnamoorthy, C. S., P. P. Venkatesh, and R. Sudarshan (2002). Objectoriented framework for genetic algorithms with application to space truss optimization. J. Computing Civil Eng. 16 (1), 66–75. Kutylowski, K. (2002). On nonunique solutions in topology optimization. Struct. Multidisc. Optim. 23, 398–403.

386

Optimization of Finite Dimensional Structures

Kwan, A. S. (1998). An evolutionary approach for layout optimization of truss structures. Int. J. Space Struct. 13 (3), 145–155. Lagaros, N., M. Papadrakakis, and G. Kokossalakis (2002). Structural optimization using evolutionary algorithm. Comput. Struct. 80, 571–598. Lagaros, N., L. D. Psarras, M. Papadrakakis, and G. Kokossalakis (2008). Optimum design of steel structures with web opening. Eng. Struct. 30, 2528–2537. Laguna, M. and R. Marti (2003). Scatter-Search Methodology and Implementation in C. Dordrecht, The Netherlands: Kluwer Academic Publishers. Lamberti, L. (2008). An eﬃcient simulated annealing algorithm for design optimization of truss structures. Comput. Struct. 86, 1936–1953. Lee, K. S. and Z. W. Geem (2004). A new structural optimization method based on the harmony search. Comput. Struct. 82, 781–798. Lee, T. H. and J. J. Jung (2007). Kriging metamodel based optimization. In J. S. Arora (Ed.), Optimization of Structural and Mechanical Systems. Singapore: World Scientiﬁc. Leelataviwat, S., S. C. Goel, and B. Stojadinovi´c (2002). Energy-based seismic design of structures using yield mechanism and target drift. J. Struct. Eng. 128 (8), 1046–1054. Lemonge, A. C. C., H. J. C. Barbosa, and L. G. Fonseca (2009). A genetic algorithm for conﬁguration and sizing optimization of dome structures including cardinality constraints. In Proc. 8th World Congress of Structural and Multidisciplinary Optimization (WCSMO8), Paper No. 1409, Lisbon. Lewi´ nski, T. and G. I. N. Rozvany (2007). Exact analytical solutions for some popular benchmark problems in topology optimization, II: Threesided polygonal supports. Struct. Multidisc. Optim. 33, 337–349. Li, C., R. Priemer, and K.-H. Cheng (2004). Optimization by random search with jumps. Int. J. Numer. Methods Eng. 60, 1301–1315. Li, G., R.-G. Zhou, L. Duan, and W.-F. Chen (1999). Multiobjective and multilevel optimization for steel frames. Eng. Struct. 21, 519–529. Li, L. J., Z. B. Huang, F. Liu, and Q. H. Wu (2007). A heuristic particle swarm optimizer for optimization of pin connected structures. Comput. Struct. 85, 340–349. Liew, J. Y. R., D. W. White, W. F. Chen, and S. Toma (1993a). Secondorder reﬁned plastic hinge analysis for frame design, Part I. J. Struct. Eng. 119 (3), 3196–3216. Liew, J. Y. R., D. W. White, W. F. Chen, and S. Toma (1993b). Secondorder reﬁned plastic hinge analysis for frame design, Part II. J. Struct. Eng. 119 (11), 3217–3237.

References

387

Lin, J. H., W. Y. Che, and Y. S. Yu (1982). Structural optimization on geometrical conﬁguration and element sizing with statical and dynamical constraints. Comput. Struct. 15, 507–515. Lio, M. D., V. Cossalter, and R. Lot (2000). On the use of natural coordinates in optimal synthesis of mechanisms. Mechanism and Machine Theory 35, 1367–1389. Liu, M. and S. A. Burns (2003). Multiple fully stressed designs of steel frame structures with semi-rigid connections. Int. J. Numer. Methods Eng. 58, 821–838. Liu, M., S. A. Burns, and Y. K. Wen (2006). Genetic algorithm based construction-conscious minimum weight design of seismic steel momentresisting frames. J. Struct. Eng. 132 (1), 50–58. Liu, M., Y. K. Wen, and S. A. Burns (2004). Life cycle cost oriented seismic design optimization of steel moment frame structures with risk-taking preference. Eng. Struct. 26, 1407–1421. Liu, M., Y. K. Wen, and S. A. Burns (2005). Multiobjective optimization for performance-based seismic design of steel moment-resisting frames structures. Earthquake Eng. Struct. Dyn. 34, 289–306. Liu, S. and H. Qiao (2009). Topology optimization of continuum structures with diﬀerent tensile and compressive properties in bridge layout design. In Proc. 8th World Congress of Structural and Multidisciplinary Optimization (WCSMO8), Paper No. 1328, Lisbon. Luco, J. E. and H. L. Wong (1986). Response of a rigid foundation to a spatially random ground motion. Earthquake Eng. Struct. Dyn. 14, 891–908. Luenberger, D. G. (2003). Linear and Nonlinear Programming. Boston: Kluwer Academic Publishers. Luh, G.-C. and C.-H. Chueh (2004). Multi-objective optimal design of truss structure with immune algorithm. Comput. Struct. 82, 829–844. Luh, G.-C. and C.-Y. Lin (2008). Optimal design of truss structures using ant algorithm. Struct. Multidisc. Optim. 36, 365–379. Luo, Z.-Q., J.-S. Pang, and D. Ralph (1996). Mathematical Programs with Equilibrium Constraints. Cambridge, UK: Cambridge University Press. Maar, B. and V. Schulz (2000). Interior point multigrid methods for topology optimization. Struct. Multidisc. Optim. 19, 214–224. Machaly, E. S. B. (1986). Optimum weight analysis of steel frames with semirigid connections. Comput. Struct. 23 (4), 461–474. Mahin, S., J. Malley, and R. Hamburger (2002). Overview of the FEMA/SAC program for reduction of earthquake hazards in steel moment frame structures. J. Const. Steel Res. 58, 511–528.

388

Optimization of Finite Dimensional Structures

Makode, P. V., R. B. Corotis, and M. R. Ramirez (1999). Nonlinear analysis of frame structures by pseudodistortions. J. Struct. Eng. 125 (11), 1309– 1317. Mangasarian, O. L. (1969). Nonlinear Programming. New York: McGrawHill. Marcelin, J. L., P. Trompette, and R. Dornberger (1995). Optimization of composite beam structures using a genetic algorithm. Struct. Opt. 9, 236–244. Marler, T. and J. S. Arora (2004). Survey of multi-objective optimization methods for engineering. Struct. Multidisc. Optim. 26 (6), 369–395. Mart´ınez, P., P. Mart´ı, and O. M. Querin (2007). Growth method for size, topology, and geometry optimization of truss structures. Struct. Multidisc. Optim. 33, 13–26. Massa, F., B. Lallemanda, and T. Tison (2009). Fuzzy multiobjective optimization of mechanical structures. Comput. Methods Appl. Mech. Engrg. 198, 631–643. Masur, E. F. (1984). Optimal structural design under multiple eigenvalue constraints. Int. J. Solids Struct. 20, 211–231. Maxwell, J. C. (1890). On reciprocal ﬁgures, frames, and diagrams of forces. Scientiﬁc Papers 2, 161–207. McKeown, J. J. (1998). Growing optimal pin-jointed frames. Struct. Opt. 15, 92–100. McNeil, W. A. (1971). Structural weight minimization using necessary and suﬃcient conditions. J. Optimization Theory Appl. 8 (6), 454–466. Mehrotra, S. (1992). On the implementation of a primal-dual interior point method. SIAM J. Optim. 2, 575–601. Meirovitchand, L. and A. L. Hale (1981). On the substructure synthesis method. AIAA J. 19 (7), 940–947. Melosh, R. J. and R. Luik (1968). Multiple conﬁguration analysis of structures. J. Struct. Div. ASCE 94 (ST11), 2581–2596. Mesarovi´c, M. D., D. Macko, and Y. Takahara (1970). Theory of Hierarchical, Multilevel, Systems. New York: Academic Press. Metropolis, N., A. Rosenbluth, M. Rosenbluth, A. Teller, and E. Teller (1953). Equation of state calculations by fast computing machines. J. Chem. Phys 21, 1087–1092. Michell, A. G. M. (1904). The limits of economy in frame structures. Philosophical Magazine Sect. 6, 8 (47), 589–597. Mijar, A. R., C. Swan, J. S. Arora, and I. Kosaka (1998). Continuum topology optimization for concept design of frame bracing system. J. Struct. Eng. 124 (5), 541–550.

References

389

Mittelmann, H. D. and D. Roose (Eds.) (1990). Continuation Techniques and Bifurcation Problems. Basel, Germany: Birkh¨auser Verlag. Mohammadi, R. K., M. H. El Naggar, and H. Moghaddam (2004). Optimum strength distribution for seismic resistant shear buildings. Int. J. Solids Struct. 41, 6597–6612. Mohan, C. and H. T. Nguyen (1999). A controlled random search technique incorporating the simulated annealing concept for solving integer and mixed integer global optimization problems. Comp. Opt. Appl. 14, 103– 132. Mr´ oz, Z. and D. Bojczuk (2003). Finite topology variations in optimal design of structures. Struct. Multidisc. Optim. 25, 153–173. MSC Software (2005). ADAMS 2005 User’s Manual. Santa Ana, CA: MSC Software Corp. Mueller, K. M., M. Liu, and S. A. Burns (2002). Fully stressed design of frame structures and multiple load paths. J. Struct. Eng. 128 (6), 806– 814. Munro, J. and P.-H. Chuang (1986). Optimal plastic design with imprecise data. J. Eng. Mech. Div. ASCE 112 (9), 888–903. Muralidhar, R. and J. Rao (1997). New models for optimal truss topology in limit design based on uniﬁed elastic/plastic analysis. Comput. Methods Appl. Mech. Engrg. 140, 109–138. Myers, R. H. and D. C. Montgomery (1995). Response Surface Methodology: Process and Product Optimization using Design Experiments. New York: Wiley. Nagtegaal, J. C. (1973). A new approach to optimal design of elastic structures. Comput. Methods Appl. Mech. Engrg. 2, 255–264. Nakamura, T. and M. Ohsaki (1988). Sequential optimal truss generator for frequency ranges. Comput. Methods Appl. Mech. Engrg. 67 (2), 189–209. Nakamura, T. and M. Ohsaki (1992). A natural generator of optimum topology of plane trusses for speciﬁed fundamental frequency. Comput. Methods Appl. Mech. Engrg. 94 (1), 113–129. Nakanishi, Y. and S. Nakagiri (1996a). Optimization of frame topology using boundary cycle and genetic algorithm. JSME Int. J., Series A. 39 (2), 279–285. Nakanishi, Y. and S. Nakagiri (1996b). Optimization of truss topology using boundary cycle: Derivation of design variables to avoid inexpedient structure. JSME Int. J., Series A. 39 (3), 415–421. Nakanishi, Y. and S. Nakagiri (1997). Structural optimization under topological constraint represented by hom*ology groups: Topological constraint on one-dimensional complex by use of zero- and one-dimensional groups. JSME Int. J., Series A. 40 (3), 219–227.

390

Optimization of Finite Dimensional Structures

Nakayama, H. (1995). Aspiration level approach to interactive multiobjective programming and its applications. In P. M. Pardalos, Y. Siskos, and C. Zopounidis (Eds.), Advances in Multicriteria Analysis, Dordrecht, The Netherlands, pp. 147–174. Kluwer Academic Publishers. Narayanan, S. (2006). Space Structures: Principles and Practice, Vol. 1,2. Brentwood, UK: Multi-Science Publishing. Newmark, N. M. and W. J. Hall (1982). Earthquake spectra and design. Technical Report, Earthquake Engineering Research Institute, Berkeley, CA. Nguyen, D. T. (1987). Multilevel substructuring sensitivity analysis. Comput. Struct. 25 (2), 191–202. Nguyen, V., J. Strodiot, and C. Fleury (1987). A mathematical convergence analysis of the convex linearization method for engineering design. Eng. Opt. 11, 195–216. Nishiwaki, S., S. Min, J. Yoo, and N. Kikuchi (2001). Optimal structural design considering ﬂexibility. Comput. Methods Appl. Mech. Engrg. 190, 4457–4504. Noor, A. K. and A. E. Lowder (1975). Structural reanalysis via a mixed method. Comput. Struct. 5, 9–12. Oberndorfer, J. M., W. Achtziger, and H. R. E. M. H¨ornlein (1996). Two approaches for truss topology optimization: A comparison for practical use. Struct. Opt. 11, 137–144. Ogawa, K. and M. Tada (1994). Computer program for static and dynamic analysis of steel frames considering the deformation of joint panel. In Proc. 17th Symp. on Computational Technology of Information, System and Applications, pp. 79–84 (in Japanese). Ohsaki, M. (1995). Genetic algorithm for topology optimization of trusses. Comput. Struct. 57 (2), 219–225. Ohsaki, M. (1997a). Optimization of building structural systems using parametric optimization method. J. Struct. Eng., Architectural Inst. Japan 43B, 79–88 (in Japanese). Ohsaki, M. (1997b). Simultaneous optimization of topology and geometry of a regular plane truss. Comput. Struct. 66 (1), 69–77. Ohsaki, M. (2001a). Random search method based on exact reanalysis for topology optimization of trusses with discrete cross-sectional areas. Comput. Struct. 79 (6), 673–679. Ohsaki, M. (2001b). Sensitivity of optimum designs for spatially varying ground motions. J. Struct. Eng. 127 (11), 1324–1329. Ohsaki, M. (2003a). Application of optimization methods to architectural structures. In G. Yagawa (Ed.), Handbook of Structural Engineering, pp. 647–649. Tokyo: Maruzen (in Japanese).

References

391

Ohsaki, M. (2003b). Earthquake response analysis and optimum design of arch-type trusses considering stiﬀness of supporting structure. J. Struct. Constr. Eng., Architectural Inst. Japan (566), 53–58 (in Japanese). Ohsaki, M. (2003c). System engineering. In J. Kanda (Ed.), Architecture and Engineering, Visual Introduction of Architecture 9, pp. 204–211. Tokyo: Shokokusha (in Japanese). Ohsaki, M. (2005a). Design sensitivity analysis and optimization for nonlinear buckling of ﬁnite-dimensional elastic conservative structures. Comput. Methods Appl. Mech. Engrg. 194, 3331–3358. Ohsaki, M. (2005b). Spatial structures: Optimization of structures with discrete cross-sectional areas using heuristic approaches. In Intelligible Soft Computing for Architecture, City and Environment, pp. 77–82. Architectural Institute of Japan (in Japanese). Ohsaki, M. (2006a). Local and global searches of approximate optimal designs of regular frames. Int. J. Numer. Methods Eng. 67, 132–147. Ohsaki, M. (2006b). Local and global searches of approximate optimal designs of regular frames. In Proc. 3rd European Conference on Computational Mechanics (ECCM 2006), Paper No. 1174, Lisbon. Ohsaki, M. (2008). Local search for multiobjective optimization of steel frames. In Proc. 5th China-Japan-Korea Joint Symp. on Optimization of Structural and Mechanical Systems (CJK-OSM5), Jeju, Korea. Ohsaki, M. and J. S. Arora (1993). A direct application of higher order parametric programming techniques to structural optimization. Int. J. Numer. Methods Eng. 36, 2683–2702. Ohsaki, M. and J. S. Arora (1994). Design sensitivity analysis of elastoplastic structures. Int. J. Numer. Methods Eng. 37, 737–762. Ohsaki, M., K. Fujisawa, N. Katoh, and Y. Kanno (1999). Semi-deﬁnite programming for topology optimization of trusses under multiple eigenvalue constraints. Comput. Methods Appl. Mech. Engrg. 180, 203–217. Ohsaki, M. and M. Hayashi (2000). Fairness metrics for shape optimization of ribbed shells. J. Int. Assoc. Shells and Spatial Struct. 41 (1), 31–39. Ohsaki, M. and K. Ikeda (2007). Stability and Optimization of Structures – Generalized Sensitivity Analysis. Mechanical Engineering Series. New York: Springer. Ohsaki, M. and Y. Kanno (2007). Semideﬁnite programming for engineering applications. In J. S. Arora (Ed.), Optimization of Structural and Mechanical Systems, pp. 541–567. Singapore: World Scientiﬁc. Ohsaki, M. and Y. Kato (1997). Simultaneous optimization of topology and nodal locations of a plane truss associated with a B´ezier curve. In Proc. Annual Meeting, Architectural Inst. Japan, Volume B-1, pp. 447–448 (in Japanese).

392

Optimization of Finite Dimensional Structures

Ohsaki, M. and Y. Kato (1999). Simultaneous optimization of topology and nodal locations of a plane truss associated with a B´ezier curve. In Structural Engineering in the 21st Century, Proc. Structures Congress, pp. 582–585. ASCE. Ohsaki, M. and N. Katoh (2005). Topology optimization of trusses with stress and local constraints on nodal stability and member intersection. Struct. Multidisc. Optim. 29, 190–197. Ohsaki, M., N. Katoh, T. Kinosh*ta, S. Tanigawa, D. Avis, and I. Streinu (2009). Enumeration of optimal pin-jointed bistable compliant mechanisms with non-crossing members. Struct. Multidisc. Optim. 37, 645– 651. Ohsaki, M., Y. Nagano, and K. Wakamatsu (2000). A two-level optimization method for seismic design of elastic three-dimensional frames. In Proc. 12th World Conference on Earthquake Engineering, Paper #0686, Auckland, New Zealand. Ohsaki, M. and T. Nakamura (1993). A natural generator of optimum topology of a space frame for speciﬁed fundamental frequency. In G. Parke and C. Howard (Eds.), Space Structures 4, Proc. 4th Int. Conf. on Space Structures, pp. 1221–1229. Guildford, UK: Thomas Telford. Ohsaki, M. and T. Nakamura (1994). Optimum design with imperfection sensitivity coeﬃcients for limit point loads. Struct. Opt. 8, 131–137. Ohsaki, M. and T. Nakamura (1996). Minimum constraint perturbation method for topology optimization of systems. Eng. Opt. 26, 171–186. Ohsaki, M., T. Nakamura, and Y. Isshiki (1998). Shape-size optimization of plane trusses with designer’s preference. J. Struct. Eng. 124 (11), 1323– 1330. Ohsaki, M., T. Nakamura, and M. Kohiyama (1997). Shape optimization of a double-layer space truss described by a parametric surface. Int. J. Space Struct. 12 (2), 109–119. Ohsaki, M. and S. Nishiwaki (2004). Shape design of bistable compliant mechanism by utilizing snapthrough behavior. Trans. Japan Soc. of Mech. Eng., Ser. A 70 (700), 23–28 (in Japanese). Ohsaki, M. and S. Nishiwaki (2005). Shape design of pin-jointed multistable compliant mechanism using snapthrough behavior. Struct. Opt. 30, 327– 334. Ohsaki, M. and S. Nishiwaki (2007a). Generation of link mechanism by shape-topology optimization of trusses considering geometrical nonlinearity. In Proc. 7th World Congress of Structural and Multidisciplinary Optimization (WCSMO7), Seoul. (also available in J. Computational Science and Technology, JSME 3 (1), pp. 46–53, 2009).

References

393

Ohsaki, M. and S. Nishiwaki (2007b). Generation of link mechanism by shape-topology optimization of trusses considering geometrical nonlinearity. Trans. Japan Soc. of Mech. Eng., Ser. A 73 (729), 659–665 (in Japanese). Ohsaki, M., T. Ogawa, and R. Tateishi (2003). Shape optimization of curves and surfaces considering fairness metrics and elastic stiﬀness. Struct. Multidisc. Optim. 24, 449–456. Erratum: 27, pp. 250–258, 2004. Ohsaki, M., H. Tagawa, and Y. Kato (2000). Optimum design of structures subjected to spatially varying ground motion. J. Struct. Eng., Architectural Inst. Japan 46B, 9–18 (in Japanese). Ohsaki, M., H. Tagawa, and P. Pan (2009). Shape optimization of reduced beam section for maximum plastic energy dissipation under cyclic loads. J. Const. Steel Res. 65, 1511–1519. Ohsaki, M. and R. Watada (2008). Linear mixed integer programming for topology optimization of trusses and plates. In Proc. 6th Int. Conf. on Computation of Shell and Spatial Structures, IASS-IACM, Ithaca, NY. Ohsaki, M., J. Y. Zhang, and S. Kimura (2005). An optimization approach to design of geometry and forces of tensegrities. In Proc. IASS Symp. 2005, Bucharest, pp. 603–610. Int. Assoc. Shell and Spatial Struct. Ohsaki, M., J. Y. Zhang, and Y. Ohishi (2008). Force design of tensegrity structures by enumeration of vertices of feasible region. Int. J. Space Struct. 23 (2), 117–126. Olhoﬀ, N. (1980). Optimal design with respect to structural eigenvalues. In Proc. 15th IUTAM Congress, Toronto, Canada, pp. 133–149. Olhoﬀ, N. and S. H. Rasmussen (1977). On single and bimodal optimum buckling loads of clamped columns. Int. J. Solids Struct. 13, 605–614. Olhoﬀ, N. and J. E. Taylor (1979). On optimal structural remodelling. J. Optimization Theory Appl. 27 (4), 571–582. Orozco, C. and O. Ghattas (1997). A reduced SAND method for optimal design of non-linear structures. Int. J. Numer. Methods Eng. 40, 2759– 2774. Otto, F. (1967). Tensile Structures: Volume 1, Pneumatic Structures. Cambridge, MA: MIT Press. Otto, F. (1969). Tensile Structures: Volume 2, Cable Structures. Cambridge, MA: MIT Press. ¨ Ozakca, M., E. Hinton, and N. V. R. Rao (1993). Shape optimization of axisymmetric structures with adaptive ﬁnite element procedures. Struct. Opt. 5, 256–264. Paeng, J. K. and J. S. Arora (1989). Dynamic response optimization of mechanical systems with multiplier methods. AIAA J. 111, 73–80.

394

Optimization of Finite Dimensional Structures

Pan, P., M. Ohsaki, and T. Kinosh*ta (2007). Constraint approach to performance-based design of steel moment-resisting frames. Eng. Struct. 29, 186–194. Pan, P., M. Ohsaki, and H. Tagawa (2007). Shape optimization of H-beam ﬂange for maximum plastic energy dissipation. J. Struct. Eng. 133 (8), 1176–1179. Papalambos, P. Y. and D. J. Wilde (2000). Principles of Optimal Design: Modeling and Computation (2nd ed.). Cambridge, UK: Cambridge University Press. Patnaik, S. N. and P. Dayaratnam (1970). Behavior and design of pin connected structures. Int. J. Numer. Methods Eng. 2, 579–595. Patnaik, S. N. and A. Hopkins (1998). Optimality of a fully stressed design. Comput. Methods Appl. Mech. Engrg. 165, 215–221. Pavlovˇciˇc, L., A. Krajnc, and D. Beg (2004). Cost function analysis in the structural optimization of steel frames. Struct. Multidisc. Optim. 28, 286–295. Pedersen, P. (1972). On the optimal layout of multi-purpose trusses. Comput. Struct. 2, 695–712. Pedersen, P. (1973). Optimal joint positions for space trusses. J. Struct. Div. ASCE 99 (ST12), 2459–2476. Pedersen, P. and N. L. Pedersen (2009). A discussion on the application of optimality criteria for compliance with forced support displacement. In Proc. 8th World Congress of Structural and Multidisciplinary Optimization (WCSMO8), Paper No. 1047, Lisbon. Pellegrino, S. (1993). Structural computation with the singular value decomposition of the equilibrium matrix. Int. J. Solids Struct. 30 (21), 3025–3035. Pellegrino, S. and C. R. Calladine (1986). Matrix analysis of statically and kinematically indeterminate frameworks. Int. J. Solids Struct. 22 (4), 409–428. Peressini, A. L., F. E. Sullivan, and J. J. Uhl (1988). The Mathematics of Nonlinear Programming. New York: Springer. Pereyra, V., D. Lawver, and J. Isenberg (2003). An algorithm for optimal design of steel frame structures. Applied Numerical Mathematics 47, 503–514. Petersson, J. (1999). A ﬁnite element analysis of optimal variable thickness sheets. SIAM J. Numer. Anal. 36 (6), 1759–1778. Petersson, J. (2001). On continuity of the design-to-state mappings for trusses with variable topology. J. Eng. Sci. 39, 1119–1141.

References

395

Pierre, D. A. and M. J. Lowe (1975). Mathematical Programming via Augmented Lagrangians. Reading, MA: Addison-Wesley. P´ olik, I. (2005). Addendum to the SeDuMi User Guide, Ver. 1.1. Advanced Optimization Laboratory, McMaster University, Ontario, Canada. Available at: http://sedumi.ie.lehigh.edu/. Prager, W. (1967). Optimum plastic design of a portal frame for alternative loads. J. Appl. Mech. 34, 772–773. Prager, W. (1971). Foulkes mechanism in optimal plastic design for alternative loads. Int. J. Mech. Sci. 13, 971–973. Prager, W. (1972). Conditions for structural optimality. Comput. Struct. 2, 833–840. Prager, W. (1974a). Introduction to Structural Optimization. Vienna: Springer. Prager, W. (1974b). A note on discretized Michell structure. Comput. Methods Appl. Mech. Engrg. 3, 349–355. Prager, W. (1976). Geometric discussion of the optimal design of a simple truss. J. Struct. Mech. 4 (1), 57–63. Prager, W. and J. E. Taylor (1968). Problem of optimal structural design. J. Appl. Mech. 35 (1), 102–106. Price, T. E. and M. O. Eberhard (1998). Eﬀects of spatially varying ground motions on short bridges. J. Struct. Eng. 124 (8), 948–955. Price, W. L. (1983). Global optimization by controlled random search. J. Optimization Theory Appl. 40, 333–348. Pugnale, A. and M. Sassone (2007). Morphogenesis and structural optimization of shell structures with the aid of a genetic algorithm. J. Int. Assoc. Shells and Spatial Struct. 48 (3), 161–166. Putresza, J. T. and P. Kolakowski (2001). Sensitivity analysis of frame structures: Virtual distortion method approach. Int. J. Numer. Methods Eng. 50, 1307–1329. Qiu, G. Y. and X. S. Li (2010). A note on the derivation of global stress constraints. Struct. Multidisc. Optim. 40, 625–628. Rahmatalla, S. and C. Swan (2003). Form ﬁnding of sparse structures with continuum topology optimization. J. Struct. Eng. 129 (12), 1707–1716. Rajasekaran, S., V. S. Mohan, and O. Khamis (2004). The optimization of space structures using evolution strategies with functional networks. Engineering with Computers 20, 75–87. Rajeev, S. and C. S. Krishnamoorthy (1997). Genetic algorithms-based methodologies for design optimization of trusses. J. Struct. Eng. 123, 350–358.

396

Optimization of Finite Dimensional Structures

Ramm, E., K.-U. Bletzinger, and R. Reitinger (1993). Shape optimization of shell structures. Bulletin of Int. Assoc. for Shell and Spatial Struct 34 (2), 103–121. Ramrakhyami, D. S., M. I. Frecker, and G. A. Lesieutre (2009). Hinged beam elements for the topology design of compliant mechanisms using the ground structure approach. Struct. Multidisc. Optim. 37, 557–567. Rando, T. and J. A. Roulier (1991). Designing faired parametric surfaces. Comput.-Aided Des. 23, 492–497. Rao, S. S. (1987). Multi-objective optimization of fuzzy structural systems. Int. J. Numer. Methods Eng. 24, 1157–1171. Ray, T. and K. M. Liew (2002). A swarm metaphor for multiobjective design optimization. Eng. Opt. 34 (2), 141–153. Razani, R. (1965). Behavior of fully stressed design of structures and its relationship to minimum weight design. AIAA J. 3, 2262–2268. Rechenberg, I. (1965). Cybernetic solution path of an experimental problem. Library Translation 1122, Royal Aircraft Establishment, Franborough, UK. Reddy, G. M. and J. Cagan (1995a). An improved shape annealing algorithm for truss topology generation. J. Mech. Eng. 117, 315–321. Reddy, G. M. and J. Cagan (1995b). Optimally directed truss topology generation using shape annealing. J. Mech. Eng. 117, 206–209. Reemtsen, R. and J.-J. R¨ uckmann (Eds.) (1998). Semi-Inﬁnite Programming. Norwell, MA: Kluwer Academic Publishers. Reeves, C. (1995). Modern Heuristic Techniques for Combinatorial Problems. New York: McGraw-Hill. Richtmyer, R. D. and K. W. Morton (1967). Diﬀerence Methods for Initial Value Problems. New York: Interscience. Ringertz, U. T. (1985). On topology optimization of trusses. Eng. Opt. 9, 209–218. Ringertz, U. T. (1986). A branch and bound algorithm for topology optimization of truss structures. Eng. Opt. 10, 111–124. Rocakfellar, R. T. (1970). Convex Analysis. Princeton, NJ: Princeton University Press. Rodorigues, H. C., J. M. Guedes, and M. P. Bendsøe (1995). Necessary conditions for optimal design of structures with a nonsmooth eigenvalue based criterion. Struct. Opt. 9, 52–56. Rogers, D. F. and J. A. Adams (1990). Mathematical Elements for Computer Graphics. New York: McGraw-Hill.

References

397

Roulier, J. A. and T. Rando (1994). Measures of fairness for curves and surfaces. In N. S. Spadis (Ed.), Designing Fair Curves and Surfaces, pp. 75–122. Philadelphia, PA: SIAM. Rozvany, G. I. N. (1976). Optimal Design of Flexural Systems. Oxford, UK: Pergamon Press. Rozvany, G. I. N. (1989). Structural Design via Optimality Criteria. Dordrecht, The Netherlands: Kluwer Academic Publishers. Rozvany, G. I. N. (1996). Diﬃculties in truss topology optimization with stress, local buckling and system stability constraints. Struct. Opt. 11, 213–217. Rozvany, G. I. N. (Ed.) (1997). Topology Optimization in Structural Mechanics. CISM Courses and Lectures 374. Vienna: Springer-Verlag. Rozvany, G. I. N. (2001). On design-dependent constraints and singular topologies. Struct. Multidisc. Optim. 21 (2), 164–172. Rozvany, G. I. N. and M. Zhou (1994). Optimality criteria methods for large discretized systems. In H. Adeli (Ed.), Advances in Design Optimization, pp. 41–108. London: Chapman & Hall. Rozvany, G. I. N., M. Zhou, and T. Birker (1992). Generalized shape optimization without hom*ogenization. Struct. Optim. 4, 250–252. Rule, W. K. (1994). Automatic truss design by optimized growth. J. Struct. Eng. 120 (10), 3063–3070. Rychter, Z. and A. Musiuk (2007). Topological sensitivity to diagonal member ﬂips of two-layered statically determinate trusses under worst loading. Int. J. Solids Struct. 44, 4942–4957. Sadek, E. A. (1986). Dynamic optimization of framed structures with variable layout. Int. J. Numer. Methods Eng. 23, 1273–1294. Sadek, E. A. (1989). An optimality criterion method for dynamic optimization of structures. Int. J. Numer. Methods Eng. 28, 579–592. Sadek, E. A. (1992). Optimization of structures having general crosssectional relationships using an optimality criterion method. Comput. Struct. 43 (5), 959–969. Saka, M. P. (1980). Shape optimization of trusses. J. Struct. Div. ASCE 106 (ST5), 1155–1174. Saka, M. P. (2007). Optimum topological design of geometrically nonlinear single layer latticed domes using coupled genetic algorithm. Comput. Struct. 85, 1635–1646. Saka, M. P. and F. Erdal (2009). Harmony search based algorithm for the optimum design of grillage systems to LRFD-AISC. Struct. Multidisc. Optim. 38, 25–41.

398

Optimization of Finite Dimensional Structures

Saka, M. P. and M. Ulker (1991). Optimum design of geometrically nonlinear space trusses. Comput. Struct. 42 (3), 289–299. Salajegheh, E. (1996). Approximate discrete variable optimization of frame structures with dual method. Int. J. Numer. Methods Eng. 39, 1607– 1617. Salerno, G. (1992). How to recognize the order of inﬁnitesimal mechanisms: A numerical approach. Int. J. Numer. Methods Eng. 35, 1351–1395. Sankaranarayanan, S., R. T. Haftka, and R. K. Kapania (1994). Truss topology optimization with simultaneous analysis and design. AIAA J. 32 (2), 420–424. Sarma, K. and H. Adeli (2002). Life-cycle cost optimization of steel structures. Int. J. Numer. Methods Eng. 55, 1451–1462. Save, M. A. (1983). Remarks on minimum-volume designs of a three-bar truss. J. Struct. Mech. 11 (1), 101–110. Sawada, K. and A. Matsuo (2003). A revised enumeration algorithm for elastic plastic discrete optimization of steel building frames. J. Struct. Constr. Eng., Architectural Inst. Japan 574, 93–98 (in Japanese). Saxena, A. (2005). Synthesis of compliant mechanisms for path generation using genetic algorithm. J. Mech. Design 127, 745–752. Saxena, A. and G. K. Ananthasuresh (2000). On an optimal property of compliant topologies. Struct. Multidisc. Optim. 19, 36–49. Schmidt, L. C. (1962). Minimum weight layouts of elastic, statically determinate, triangulated frames under alternative load systems. J. Mech. Phys. Solids 10, 139–149. Schmit, L. A. and K. J. Chang (1984). Optimum design sensitivity based on approximation concepts and dual method. Int. J. Numer. Methods Eng. 20, 39–75. Schmit, L. A. and B. Farshi (1974). Some approximation concepts for structural synthesis. AIAA J. 12, 692–699. Schmit, L. A. and R. L. Fox (1965). An integrated approach to structural synthesis and analysis. AIAA J. 3, 1104–1112. Schur, J. (1917). u ¨ber Potenzreihen, die im Innern des Einheitskreises beschr¨ ankt sind. J. Reine Angew Math. 147, 205–232. Schutte, J. F. and A. A. Groenwold (2003). Sizing design of truss structures using particle swarm. Struct. Multidisc. Optim. 25, 261–269. Sedaghati, R. and E. Esmailzadeh (2003). Optimum design of structures with stress and displacement constraints using the force method. Int. J. Mech. Sci. 45, 1369–1389.

References

399

Sekimoto, T. and H. Noguchi (2001). hom*ologous topology optimization in large displacement and buckling problems. JSME Int. J., Series A 44, 610–615. Semby, W. and A. Der Kiureghian (1985). Modal combination rules for multicomponent earthquake excitation. Earthquake Eng. Struct. Dyn. 13, 1–12. Sergeyev, O. and Z. Mr´oz (2000). Sensitivity analysis and optimal design of 3D frame structures for stress and frequency constraints. Comput. Struct. 75, 167–185. Sewell, M. J. (1987). Maximum and Minimum Principles. Cambridge, UK: Cambridge University Press. Seyranian, A. P. (1993). Sensitivity analysis of multiple eigenvalues. Mech. Struct. & Mach. 21, 261–284. Seyranian, A. P., E. Lund, and N. Olhoﬀ (1994). Multiple eigenvalues in structural optimization problem. Struct. Opt. 8, 207–227. Shames, I. H. and F. A. Cozzarelli (1997). Elastic and Inelastic Stress Analysis. Washington, DC: Taylor & Francis. Shea, K., J. Cagan, and S. J. Fenves (1997). A shape annealing approach to optimal truss design with dynamic grouping of members. J. Mech. Eng. 118, 388–394. Shea, K. and F. C. Smith (2006). Improving full-scale transmission tower design through topology and shape optimization. J. Struct. Eng. 132 (5), 781–790. Sherman, J. and W. J. Morrison (1950). Adjustment of an inverse matrix corresponding to changes in one element of a given matrix. Ann. Math. Statist. 21, 124–127. Sheu, C. Y. and L. A. Schmit (1972). Minimum weight design of elastic redundant trusses under multiple static loading conditions. AIAA J. 10 (2), 155–162. Shield, R. T. (1960). On the optimum design of shells. J. Appl. Mech. 27, 316–322. Shimizu, K., Y. Ishizuka, and J. F. Bard (1997). Nondiﬀerentiable and Two-Level Mathematical Programming. Norwell, MA: Kluwer Academic Publishers. Shin, Y. S., R. T. Haftka, and R. H. Plaut (1987). Simultaneous analysis and design or eigenvalue maximization. AIAA J. 26 (6), 738–744. Shin, Y. S., R. T. Haftka, L. T. Watson, and R. H. Plaut (1988). Tracing structural optima as a function of available resources by a hom*otopy method. Comput. Methods Appl. Mech. Engrg. 70, 151–164.

400

Optimization of Finite Dimensional Structures

Sim˜ oes, L. M. C. (1989). Isolated global optimality in truss sizing problems. Comput. Struct. 33 (2), 375–384. Smith, B., P. Bjørstad, and W. Gropp (1996). Domain Decomposition: Parallel Multilevel Methods for Elliptic Partial Diﬀerential Equations. New York: Cambridge University Press. Sobieszczanski-Sobieski, J. (1992). A technique for locating function roots and for satisfying equality constraints in optimization. Struct. Opt. 4 (3– 4), 241–243. Sobieszczanski-Sobieski, J. and R. T. Haftka (1996). Multidisciplinary aerospace design optimization. In Proc. 34th Aerospace Science Meeting and Exhibit, Paper AIAA-96-0711, Reno, NV. AIAA. Sobieszczanski-Sobieski, J., B. James, and A. Dovi (1985). Structural optimization by multilevel decomposition. AIAA J. 23 (11), 1775–1782. Stadler, W. (1979). A survey of multicriteria optimization of the vector maximum problem, Part I: 1776–1960. J. Optimization Theory Appl. 29 (1), 1–52. Stadler, W. (Ed.) (1988). Multicriteria Optimization in Engineering and in the Sciences. New York: Plenum Press. Stadler, W. (1999). Discrete geometry, the Steiner problem, and topology optimization. In Proc. 3rd World Congress of Structural and Multidisciplinary Optimization (WCSMO3). Stolpe, M. (2004). Global optimization of minimum weight truss topology problems with stress, displacement, and local buckling constraints using branch-and-bound. Int. J. Numer. Methods Eng. 61 (8), 1270–1309. Stolpe, M. and K. Svanberg (2001a). An alternative interpolation scheme for minimum compliance topology optimization. Struct. Multidisc. Optim. 22, 116–124. Stolpe, M. and K. Svanberg (2001b). On the trajectories of the epsilonrelaxation approach for stress-constrained truss topology optimization. Struct. Multidisc. Optim. 21, 140–151. Stolpe, M. and K. Svanberg (2003a). Modeling topology optimization problems as linear mixed 0-1 programs. Int. J. Numer. Methods Eng. 57, 723–739. Stolpe, M. and K. Svanberg (2003b). A note on stress-constrained truss topology optimization. Struct. Multidisc. Optim. 25, 62–64. Storaasli, O. O. and J. Sobieszczanski-Sobieski (1983). On the accuracy of the Taylor approximation for structural resizing. AIAA J 21, 1571–1580. Sturm, J. F. (1999). Using SeDuMi 1.02, a Matlab toolbox for optimization over symmetric cones. Optimization Methods and Software 11-12, 625– 653.

References

401

Sui, Y. K., J. Du, and Y. Guo (2006). Independent continuum mapping for topological optimization of frame structures. Struct. Multidisc. Optim. 22, 611–619. Sui, Y. K. and X. C. Wang (1997). Second-order method of generalized geometric programming for spatial frame optimization. Comput. Methods Appl. Mech. Engrg. 141, 117–123. Svanberg, K. (1981). Optimization of geometry in truss design. Comput. Methods Appl. Mech. Engrg. 28, 63–80. Svanberg, K. (1984). On local and global minima in structural optimization. In E. Atrek, R. H. Gallagher, K. M. Ragsdel, and O. C. Zienkiewicz (Eds.), New Directions in Structural Design 34, pp. 327–341. Chichester, UK: John Wiley & Sons. Svanberg, K. (1987). The method of moving asymptotes – A new method for structural optimization. Int. J. Numer. Methods Eng. 24, 359–373. Svanberg, K. (1994). On the convexity and concavity of compliances. Struct. Opt. 7, 42–46. Svanberg, K. and M. Werme (2009). On the validity of using small positive lower bounds on design variables in discrete topology optimization. Struct. Multidisc. Optim. 37, 325–334. Sved, G. (1954). The minimum weight of certain redundant structures. Austral. J. Appl. Sci. 5, 1–8. Sved, G. and Z. Ginos (1968). Structural optimization under multiple loading. Int. J. Mech. Sci. 10, 803–805. Svensson, B. (1987). A substructuring approach to optimum structural design. Comput. Struct. 25, 251–258. Tagawa, H. and M. Ohsaki (1999). A continuous topology transition model for shape optimization of plane trusses with uniform cross-sectional area. In Proc. 3rd World Congress of Structural and Multidisciplinary Optimization (WCSMO3), pp. 254–256. Takagi, J. and M. Ohsaki (2004). Design of lateral braces for columns considering critical imperfection of buckling. Int. J. Structural Stability and Dynamics 4 (1), 69–88. Taleb-Agha, G. and R. B. Nelson (1975). Method for the optimum design of truss-type structures. AIAA J. 14 (4), 436–445. Tam, T. K. H. and A. Jennings (1989). Classiﬁcation and comparison of LP formulations for the plastic design of frames. Eng. Struct. 11, 163–178. Taylor, J. E. (1967). Minimum mass bar for axial vibration at speciﬁed natural frequency. AIAA J. 5 (10), 1911–1913. Thompson, B. S. and C. K. Sung (1986). A survey of ﬁnite element techniques for mechanism design. Mechanism and Machine Theory 21 (4), 3517–359.

402

Optimization of Finite Dimensional Structures

Tonon, F. (1999). Multiobjective optimization of uncertain structures through fuzzy set and random theory. Computer-Aided Civil and Infrastructure Eng. 14, 119–140. Topping, B. H. V. (1984). Shape optimization of skeletal structures: A review. J. Struct. Eng. 109 (8), 1933–1951. Topping, B. H. V. (1992). Mathematical programming techniques for shape optimization of skeletal structures. In G. I. N. Rozvany (Ed.), Shape and Layout Optimization of Structural Systems and Optimality Criteria Methods, pp. 349–375. Vienna: Springer. Topping, B. V. H., A. I. Khan, and J. P. Leite (1996). Topological design of truss structures using simulated annealing. Struct. Engng. Rev. 8, 301–314. Tosserams, S., L. F. P. Etman, P. Y. Papalambos, and J. E. Rooda (2006). An augmented Lagrangian relaxation for analytical target cascading using the alternating direction method of multipliers. Struct. Multidisc. Optim. 31, 176–189. Tuttle, E. R., S. W. Peterson, and J. E. Titus (1989). Enumeration of basic kinematic chains using the theory of ﬁnite groups. J. Mechanisms, Transmissions, and Automation in Design, ASME 111, 498–503. Twu, S.-L. and K. K. Choi (1992). Conﬁguration design sensitivity analysis of built-up structures, Part I: Theory. Int. J. Numer. Methods Eng. 35, 1127–1150. van Keulen, F., R. T. Haftka, and N. H. Kim (2005). Review of options for structural design sensitivity analysis, Part 1: Linear systems. Comput. Methods Appl. Mech. Engrg. 194 (30-33), 3213–3243. Vanderplaats, G. N. (1988). Multidiscipline design optimization. Appl. Mech. Rev. 41, 257–262. Vanderplaats, G. N. (1999). Numerical Optimization Techniques for Engineering Design. Colorado Springs, CO: Vanderplaats Research & Development Inc. Vanderplaats, G. N. and N. Yoshida (1985). Eﬃcient calculation of optimum design sensitivity. AIAA J. 23, 1798–1803. Vassart, N., R. Laporte, and R. Motro (2000). Determination of mechanism’s order for kinematically and statically indeterminate systems. Int. J. Solids Struct. 37, 3807–3839. Venkayya, V. B. (1978). Structural optimization: A review and some recommendations. Int. J. Numer. Methods Eng. 13, 203–228. Venkayya, V. B., N. S. Khot, and L. Berke (1973). Application of optimality criteria approaches on automated design of large practical structures. In Proc. 2nd Symp. on Structural Optimization, AGARD-CP-123, Milan, Italy, pp. 3.1–3.19.

References

403

Venkayya, V. B. and V. A. Tishler (1983). Optimization of structures with frequency constraints. In Computer Methods in Nonlinear Solids Structural Mechanics, ASME-AMD-54, pp. 239–259. New York: ASME. Viana, A. and J. P. de Sousa (2000). Using metaheuristics in multiobjective resource constrained project. European J. Operational Research 120, 359–374. Villaverde, R. (1997). Seismic design of secondary structures: State of the art. J. Struct. Eng. 123 (8), 1011–1019. Visual Numerics Inc. (1997). IMSL Math/Library Ver. 4.01. Houston: Visual Numerics Inc. VR&D (1999). DOT User’s Manual, Ver 5.0. Colorado Springs, CO. Wang, Q. and J. S. Arora (2006a). Alternative formulation for structural optimization: An evaluation using frames. J. Struct. Eng. 132 (12), 1880– 1889. Wang, Q. and J. S. Arora (2006b). Optimization of large-scale truss structures using sparse SAND formulations. Int. J. Numer. Methods Eng. 69, 390–407. Watada, R. and M. Ohsaki (2009a). Continuation approach for investigation of non-uniqueness of optimal topology for minimum compliance. In Proc. 8th World Congress of Structural and Multidisciplinary Optimization (WCSMO8), Paper No. 1073, Lisbon. Watada, R. and M. Ohsaki (2009b). Topology optimization of trusses consisting of traditional layouts. In Proc. 9th Asian Paciﬁc Conf. on Shell and Spatial Struct. (APCS2009), Paper No. P0026, Nagoya, Japan. Watada, R. and M. Ohsaki (2009c). Topology optimization of trusses consisting of traditional layouts. J. Struct. Constr. Eng., Architectural Inst. Japan 74 (639), 857–863 (in Japanese). Watson, L. T. and R. T. Haftka (1989). Modern hom*otopy methods in optimization. Comput. Methods Appl. Mech. Engrg. 74 (3), 289–305. Wen, Y. K. and Y. J. Kang (2001). Minimum building life-cycle cost design criteria, I: Methodology. J. Struct. Eng. 127 (3), 330–337. Whidborne, J. F., D. W. Gu, and I. Postlethwaite (1997). Simulated annealing for multiobjective control system design. IEE Proc. Control Theory Appl. 144 (6), 582–588. Whittaker, A., M. Constantinou, and P. Tsopelas (1998). Displacement estimates for performance-based seismic design. J. Struct. Eng., ASCE 124 (8), 905–912. Wilson, E. L., A. Der Kiureghian, and E. P. Bayo (1982). A replacement of the SRSS method in seismic analysis. Earthquake Engineering and Structural Dynamics 13, 1–12.

404

Optimization of Finite Dimensional Structures

Witten, I. H. and E. Frank (2000). WEKA, Machine Learning Algorithms in Java. San Francisco, CA: Morgan Kaufmann Publishers. Wolkowicz, H., R. Saigal, and L. Vandenberghe (Eds.) (2000). Handbook of Semideﬁnite Programming: Theory, Algorithms, and Applications. Boston, MA: Kluwer Academic Publishers. Woo, T. H. and L. A. Schmit (1981). Decomposition in optimal plastic design of structures. Int. J. Solids Struct. 17, 39–56. Woodbury, M. A. (1950). Inverting modiﬁed matrices. Memorandom Report 42, Statistical Research Group, Princeton, NJ. Wrenn, G. A. (1998). An indirect method for numerical optimization using the Kreisselmeier-Steinhause function. NASA Contractor Report 4220, NASA Langley Research Center. Wu, C. C. and J. S. Arora (1987). Simultaneous analysis and design optimization of nonlinear response. Engineering with Computers 2, 53–63. Xie, Y. M. and G. P. Steven (1993). A simple evolutionary procedure for structural optimization. Comput. Struct. 49, 885–896. Xu, L. (2002). Design and optimization of semi-rigid frames structures. In S. Burns (Ed.), Recent Advancements in Optimal Structural Design, pp. 147–168. Raton, VA: ASCE Press. Xu, L., Y. Gong, and D. E. Grierson (2006). Seismic design optimization of steel building frameworks. J. Struct. Eng. 132 (2), 277–286. Xu, L. and D. E. Grierson (1993). Computer-automated design of semirigid steel frameworks. J. Struct. Eng. 199 (6), 1740–1759. Xue, Q. and C.-C. Chen (2003). Performance-based seismic design of structures: a direct displacement-based approach. Eng. Struct. 25, 1803– 1813. Yang, X. Y., Y. M. Xie, G. P. Steven, and O. M. Querin (1999a). Bidirectional evolutionary method for stiﬀness optimization. AIAA J. 37 (11), 1483–1488. Yang, X. Y., Y. M. Xie, G. P. Steven, and O. M. Querin (1999b). Topology optimization for frequencies using an evolutionary method. J. Struct. Eng. 125 (12), 1432–1438. Yang, Y. and C. K. Soh (2002). Automated optimum design of structures using genetic programming. Comput. Struct. 80, 1537–1546. Yoshikawa, N., I. Elishakoﬀ, and S. Nakagiri (1998). Worst case estimation of hom*ology design by convex analysis. Comput. Struct. 67, 191–196. Yu, M., Z.-S. Liu, and D.-J. Wang (1996). Comparison of several approximate modal methods for computing mode shape derivatives. Comput. Struct. 62 (2), 381–393.

Yun, Y. M. and B. H. Kim (2005). Optimum design of plane steel frame structures using second-order inelastic analysis and a genetic algorithm. J. Struct. Eng. 131 (12), 1820–1831. Zembaty, Z. (1996). Spatial seismic coeﬃcients, some sensitivity results. J. Eng. Mech. 122, 379–382. Zerva, A. (1990). Response of multi-span beams to spatially incoherent seismic ground motions. Earthquake Eng. Struct. Dyn. 19, 819–832. Zhou, M. (1996). Diﬃculties in truss topology optimization with stress and local buckling constraints. Struct. Opt. 11, 134–136. Zhou, M. and R. T. Haftka (1995). A comparison of optimality criteria methods for stress and displacement constraints. Comput. Methods Appl. Mech. Engrg. 124, 253–271. Zhou, M. and G. I. N. Rozvany (1991). The COC algorithm, Part II: Topological, geometrical and generalized shape optimization. Comput. Methods Appl. Mech. Engrg. 89, 309–336. Zhou, M. and G. I. N. Rozvany (1992). DCOC: A new optimality criteria method for large systems, Part I: Theory. Struct. Opt. 5, 12–25. Zhou, M. and G. I. N. Rozvany (1993). DCOC: A new optimality criteria method for large systems, Part II: Algorithm. Struct. Opt. 6, 250–262. Zou, X.-K. and C.-M. Chan (2005). An optimal resizing technique for seismic drift design of concrete buildings subjected to response spectrum and time history loadings. Comput. Struct. 83, 1689–1704.

405