Datadriven Evolutionary Optimization Integrating

Evolutionary Computation Machine Learning And

Data Science Studies In Computational
Intelligence 975 1st Ed 2021 Yaochu Jin download
https://ebookbell.com/product/datadriven-evolutionary-
optimization-integrating-evolutionary-computation-machine-
learning-and-data-science-studies-in-computational-
intelligence-975-1st-ed-2021-yaochu-jin-34293550

Explore and download more ebooks at ebookbell.com

Here are some recommended products that we believe you will be
interested in. You can click the link to download.

Datadriven Evolutionary Modeling In Materials Technology 1st Edition

Nirupam Chakraborti

https://ebookbell.com/product/datadriven-evolutionary-modeling-in-
materials-technology-1st-edition-nirupam-chakraborti-44746376

Data Driven Business Transformation How To Disrupt Innovate And Stay

Ahead Of The Competition 1st Edition Peter Jackson

https://ebookbell.com/product/data-driven-business-transformation-how-
to-disrupt-innovate-and-stay-ahead-of-the-competition-1st-edition-
peter-jackson-45083132

Data Driven Truckers Technology And The New Workplace Surveillance

Karen Levy

https://ebookbell.com/product/data-driven-truckers-technology-and-the-
new-workplace-surveillance-karen-levy-45334034

Datadriven Retailing A Nontechnical Practitioners Guide Louisphilippe

Kerkhove

https://ebookbell.com/product/datadriven-retailing-a-nontechnical-
practitioners-guide-louisphilippe-kerkhove-46490812
Datadriven Bim For Energy Efficient Building Design Saeed Banihashemi

https://ebookbell.com/product/datadriven-bim-for-energy-efficient-
building-design-saeed-banihashemi-46844598

Datadriven Approach For Biomedical And Healthcare Nilanjan Dey

https://ebookbell.com/product/datadriven-approach-for-biomedical-and-
healthcare-nilanjan-dey-47061098

Datadriven Iterative Learning Control For Discretetime Systems Ronghu

Chi

https://ebookbell.com/product/datadriven-iterative-learning-control-
for-discretetime-systems-ronghu-chi-47257398

Datadriven Fluid Mechanics Miguel A Mendez Andrea Ianiro Bernd R Noack

https://ebookbell.com/product/datadriven-fluid-mechanics-miguel-a-
mendez-andrea-ianiro-bernd-r-noack-47558366

Datadriven Science And Engineering Steven L Brunton J Nathan Kutz

https://ebookbell.com/product/datadriven-science-and-engineering-
steven-l-brunton-j-nathan-kutz-47562344
Studies in Computational Intelligence 975

Yaochu Jin
Handing Wang
Chaoli Sun

Data-Driven
Evolutionary
Optimization
Integrating Evolutionary Computation,
Machine Learning and Data Science
Studies in Computational Intelligence

Volume 975

Series Editor
Janusz Kacprzyk, Polish Academy of Sciences, Warsaw, Poland
The series “Studies in Computational Intelligence” (SCI) publishes new develop-
ments and advances in the various areas of computational intelligence—quickly and
with a high quality. The intent is to cover the theory, applications, and design
methods of computational intelligence, as embedded in the fields of engineering,
computer science, physics and life sciences, as well as the methodologies behind
them. The series contains monographs, lecture notes and edited volumes in
computational intelligence spanning the areas of neural networks, connectionist
systems, genetic algorithms, evolutionary computation, artificial intelligence,
cellular automata, self-organizing systems, soft computing, fuzzy systems, and
hybrid intelligent systems. Of particular value to both the contributors and the
readership are the short publication timeframe and the world-wide distribution,
which enable both wide and rapid dissemination of research output.
Indexed by SCOPUS, DBLP, WTI Frankfurt eG, zbMATH, SCImago.
All books published in the series are submitted for consideration in Web of Science.

More information about this series at http://www.springer.com/series/7092

Yaochu Jin · Handing Wang · Chaoli Sun

Data-Driven Evolutionary
Optimization
Integrating Evolutionary Computation,
Machine Learning and Data Science
Yaochu Jin Handing Wang
Department of Computer Science School of Artificial Intelligence
University of Surrey Xidian University
Guildford, UK Xi’an, China

Chaoli Sun
School of Computer Science
and Technology
Taiyuan University of Science
and Technology
Taiyuan, China

ISSN 1860-949X ISSN 1860-9503 (electronic)

Studies in Computational Intelligence
ISBN 978-3-030-74639-1 ISBN 978-3-030-74640-7 (eBook)
https://doi.org/10.1007/978-3-030-74640-7

© The Editor(s) (if applicable) and The Author(s), under exclusive license to Springer Nature
Switzerland AG 2021
This work is subject to copyright. All rights are solely and exclusively licensed by the Publisher, whether
the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse
of illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and
transmission or information storage and retrieval, electronic adaptation, computer software, or by similar
or dissimilar methodology now known or hereafter developed.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in this book
are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or
the editors give a warranty, expressed or implied, with respect to the material contained herein or for any
errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional
claims in published maps and institutional affiliations.

This Springer imprint is published by the registered company Springer Nature Switzerland AG
The registered company address is: Gewerbestrasse 11, 6330 Cham, Switzerland
Foreword

One of the appeals of optimization is its applicability across countless fields of

interest; another is the opportunity it affords to interface with, and benefit from,
allied disciplines.
Optimization methodologies have been transformed dramatically since I was
drawn to their allure 50 years ago. Notable advances, that I have had a particular
interest in, include those that have been made in global optimization, multi-objective
optimization, robust optimization, and black box and data-driven optimization.
These developments have greatly enhanced optimization’s capability to satisfacto-
rily address real-world problems, thereby strengthening users’ confidence. Through
global optimization, we are released from the hazards of focusing on specific, local
areas of the search space. Multi-objective optimization satisfies the dilemma of how to
balance competing multiple objectives; it offers increased transparency by providing
a family of trade-off solutions to enable domain experts to select the desired compro-
mise solution. In real-world applications, it is almost inevitable that uncertainties will
arise from either problem formulation or the implementation environment or both.
For example, these might occur through deficiencies in the fidelity of a model, the
precision of an objective or the ability to perfectly realize the proposed solution in
practice. Robust optimization methods enable practitioners to manage these uncer-
tainties. Evolutionary computing has been a catalyst that has promoted progress in
these developments.
But what if, in a design optimization, there is no analytic representation for a
particular objective or objectives? Instead, an objective (or objectives) might rely
on the outcome of computer simulations, such as dynamic systems modelling, finite
element analysis or the modelling of computational fluid dynamics. Or, the objectives
may even depend on results obtained from experiments. Further, the collected data
may be incomplete and noisy or vary in content. It will not come as a surprise,
then, to learn that evolutionary computing has an important role in dealing with such
data-driven problems as well.
Data-driven optimization forms the central focus of this textbook. Allying the
population-based metaheuristics of evolutionary computing with the use of surrogate
models and machine learning algorithms, the authors demonstrate a formidable array
of approaches to address a range of data-driven optimization challenges.
v
vi Foreword

As I started out on my adventures with optimization research (and these resulted

in a lifetime fascination), a textbook that I regularly turned to, and still do from
time to time, was Edgar and Himmelblau’s Optimization of Chemical Processes, not
because I was a chemical engineer (I was not) but because of the blend of its clarity
of exposition, its practical approach and the authors’ evident experience of applying
their knowledge to real-world applications. Data-Driven Evolutionary Optimization
has the same appeal to me; it is written by extremely experienced researchers who
have a clear view of the importance of tailoring methodologies to address practical
problems and here they provide a sound grounding in the assorted required strategies
and their underpinning methodologies.

Sheffield, UK Peter Fleming

March 2021
Preface

I started working on fitness approximation in evolutionary optimization when I moved

back to Germany from the USA in 1999 to take up a research scientist position at
the Honda Research Institute Europe in Offenbach to complete my second Ph.D.
degree. The motivation then was to use evolutionary algorithms, in particular evolu-
tion strategies to optimize the aerodynamic performance of a turbine engine by
designing the geometry of the stator or rotor blades. To accomplish this task, time-
consuming computational fluid dynamics simulations must be performed, which
prevents one from using tens of thousands of fitness evaluations, as typically done
in evolutionary optimization. To reduce the time consumption for evolutionary aero-
dynamic optimization, fitness approximation using machine learning models, also
called meta-models or surrogates, came into play. I was very much attracted by this
research topic, since it provides a nice platform to integrate evolutionary computa-
tion with neural networks, two topics I was interested in. Out of this reason, I made
much effort to promote this new area in the evolutionary computation community
and continued working on this topic whenever there was an opportunity after I moved
to the University of Surrey in 2010.
Over the past twenty years, fitness approximation in evolutionary optimization,
also known as surrogate-assisted evolutionary optimization, has grown and devel-
oped into a fascinating research field, which is now termed data-driven evolutionary
optimization. Data-driven evolutionary optimization focuses on a class of real-world
optimization problems in which no analytical mathematical functions can be estab-
lished for the objectives or constraints. Generally speaking, data-driven optimization
may include the following situations: first, simulation-based optimization, where the
quality of a solution is assessed by an iterative, computationally intensive process,
ranging from numerically solving a large set of differential equations, to training
a deep neural network on a large data set, and second, physical experiment-based
optimization, where the objective or constraint values of a candidate solution can
be evaluated only by performing physical or human experiments. This is usually
because a high-quality computer simulation of the whole system is not possible,
either because it is computationally prohibitive, e.g. numerical simulations of the
aerodynamics of the whole aircraft, or because it is intractable because the process is
not yet fully understood, e.g. the working mechanism of the human decision-making
vii
viii Preface

process. Finally, purely data-driven optimization, where only data collected in real-
life is available for optimization and neither user-designed computer simulations nor
physical experiments are allowed. For example, optimization of a complex indus-
trial process or social system. In all the above cases, the amount of the collected data
may be either small or big, and the data may be heterogeneous, noisy, erroneous,
incomplete, ill-distributed or incremental.
Clearly, data-driven evolutionary optimization involves three different but
complementary scientific disciplines, namely evolutionary computation, machine
learning and deep learning, and data science. To effectively and efficiently solve
a data-driven optimization problem, data must be properly pre-processed. Mean-
while, machine learning techniques become indispensable for handling big data,
data paucity and various degrees of uncertainty in the data. Finally, solving the opti-
mization problem becomes extremely demanding when the optimization problem is
high-dimensional or large-scale, multi-objective and time-varying.
This book aims to provide researchers, including postgraduate research students,
and industrial practitioners a comprehensive description of the state-of-the-art
methods developed for data-driven evolutionary optimization. The book is divided
into 12 chapters. For the self-containedness of the book, a brief introduction to care-
fully selected important topics and methods in optimization, evolutionary computa-
tion and machine learning is provided in Chaps. 1–4. Chapter 5 provides the funda-
mentals of data-driven optimization, including heuristics and acquisition function-
based surrogate management, followed by Chaps. 6–8, presenting ideas that use
multiple surrogates for single-objective optimization. Representative evolutionary
algorithms for solving multi- and many-objective optimization algorithms and
surrogate-assisted data-driven evolutionary multi- and many-objective optimization
are described in Chaps. 7 and 8, respectively. Approaches to high-dimensional data-
driven optimization are elaborated in Chap. 9. A plethora of techniques for transfer-
ring knowledge from unlabelled to labelled data, from cheap objectives to expensive
ones, and from cheap problems to expensive ones are presented in Chap. 10, with the
help of semi-supervised learning, transfer learning and transfer optimization. Since
data-driven optimization is a strongly application-driven research area, offline data-
driven evolutionary optimization is treated in Chap. 11, exemplified with real-world
optimization problems such as airfoil design optimization, crude oil distillation opti-
mization and trauma system optimization. Finally, deep neural architecture search as
a data-driven expensive optimization problem is highlighted in Chap. 12. Of the 12
chapters, § 3.5– 3.6, § 4.2, § 5.2, § 6.4– 6.5, § 7.2– 7.3, § 9.6– 9.7, § 11.1, §
11.3 and Chap. 12 are written by Handing Wang, and § 3.7– 3.8, § 5.4.1, § 5.5, §
6.2– 6.3, § 9.2– 9.3 and Chap. 10 by Chaoli Sun. Handing worked as Postdoctoral
Associate during 2015–18 and Chaoli at first as Academic Visitor during 2012–13
and then as Postdoctoral Associate during 2015–17 in my group at Surrey.
To make it easier for the reader to understand and use the algorithms intro-
duced in the book, the source code for most data-driven evolutionary algorithms
presented in Chaps. 5–12 is made available (http://www.soft-computing.de/DDEO/
DDEO.html) and all baseline multi-objective evolutionary algorithms introduced in
Preface ix

this book are implemented in PlatEMO, an open-source software tool for evolutionary
multi-objective optimization (https://github.com/BIMK/PlatEMO).
This book would not have been possible without the support of many previous
colleagues, collaborators and Ph.D. students of mine. First of all, I would like to
thank Prof. Dr. Bernhard Sendhoff and Prof. Dr. Markus Olhofer, with both of whom
I closely worked at the Honda Research Institute Europe during 1999–2010. After
I joined Surrey in 2010, Markus and I still maintained close collaboration on a
number of research projects on evolutionary optimization. I would also thank Prof.
Kaisa Miettinen from the University of Jyväskylä, Finland, with whom I worked
closely as Finland Distinguished Professor during 2015–17 on evolutionary multi-
objective optimization. Thanks go to Prof. Tianyou Chai and Prof. Jinliang Ding from
Northeastern University, China, with whom I also collaborate with on evolutionary
optimization as Changjiang Distinguished Professor. The following collaborators
and previous or current Ph.D. students of mine have contributed to part of the work
presented in this book: Prof. Yew-Soon Ong, Prof. Jürgen Branke, Prof. Qingfu
Zhang, Prof. Xingyi Zhang, Prof. Aimin Zhou, Prof. Ran Cheng, Prof. Xiaoyan Sun,
Dr. Ingo Paenke, Dr. Tinkle Chugh, Mr. John Doherty, Dr. Dan Guo, Dr. Cuie Yang,
Dr. Ye Tian, Dr. Cheng He, Dr. Dudy Lim, Dr. Mingh Nhgia Le, Dr. Jie Tian, Dr.
Haibo Yu, Dr. Guo Yu, Dr. Michael Hüsken, Ms. Huiting Li, Ms. Xilu Wang, Ms.
Shufen Qin, Mr. Hao Wang, Ms. Guoxia Fu, Mr. Peng Liao, Mr. Sebastian Schmitt,
Ms. Kailai Gao, Dr. Jussi Hakanen, Dr. Tatsuya Okabe, Dr. Yanan Sun, Dr. Jan
O. Jansen, Mr. Martin Heiderich, Dr. Yuanjun Huang and Dr. Tobias Rodemann.
I would also like to take this opportunity to thank Prof. Xin Yao, Prof. Gary Yen,
Prof. Kay Chen Tan, Prof. Mengjie Zhang, Prof. Richard Everson, Prof. Jonathon
Fieldsend, Prof. Dr. Stefan Kurz, Prof. Edgar Körner and Mr. Andreas Richter for
their kind support over the past two decades. Finally, financial support from EPSRC
(UK), TEKES (Finland), National Natural Science Foundation of China, Honda
Research Institute Europe, Honda R&D Europe and Bosch Germany is gratefully
acknowledged.

Guildford, UK Yaochu Jin

February 2021
Contents

1 Introduction to Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1 Definition of Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.1 Mathematical Formulation . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.1.2 Convex Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.1.3 Quasi-convex Function . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.1.4 Global and Local Optima . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2 Types of Optimization Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.2.1 Continuous Versus Discrete Optimization . . . . . . . . . . . . 5
1.2.2 Unconstrained Versus Constrained Optimization . . . . . . 7
1.2.3 Single Versus Multi-objective Optimization . . . . . . . . . . . 7
1.2.4 Deterministic Versus Stochastic Optimization . . . . . . . . . 8
1.2.5 Black-Box and Data-Driven Optimization . . . . . . . . . . . . 8
1.3 Multi-objective Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.3.1 Mathematical Formulation . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.3.2 Pareto Optimality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
1.3.3 Preference Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14
1.3.4 Preference Articulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.4 Handling Uncertainty in Optimization . . . . . . . . . . . . . . . . . . . . . . . 17
1.4.1 Noise in Evaluations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
1.4.2 Robust Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19
1.4.3 Multi-scenario Optimization . . . . . . . . . . . . . . . . . . . . . . . . 21
1.4.4 Dynamic Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
1.4.5 Robust Optimization Over Time . . . . . . . . . . . . . . . . . . . . 24
1.5 Comparison of Optimization Algorithms . . . . . . . . . . . . . . . . . . . . . 27
1.5.1 Algorithmic Efficiency . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
1.5.2 Performance Indicators . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28
1.5.3 Reliability Assessment . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34
1.5.4 Statistical Tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
1.5.5 Benchmark Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36
1.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 38

xi
xii Contents

2 Classical Optimization Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41

2.1 Unconstrained Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.1.1 The Gradient Based Method . . . . . . . . . . . . . . . . . . . . . . . . 42
2.1.2 Newton’s Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
2.1.3 Quasi-Newton Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
2.2 Constrained Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
2.2.1 Penalty and Barriers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
2.2.2 Lagrangian Multipliers . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
2.3 Derivative-Free Search Methods . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
2.3.1 Line Search and Pattern Search . . . . . . . . . . . . . . . . . . . . . 47
2.3.2 Nelder-Mead Simplex Method . . . . . . . . . . . . . . . . . . . . . . 47
2.3.3 Model-Based Derivative-Free Search Methods . . . . . . . . 48
2.4 Deterministic Global Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . 49
2.4.1 Lipschitzian-Based Methods . . . . . . . . . . . . . . . . . . . . . . . 49
2.4.2 DIRECT . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
2.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
3 Evolutionary and Swarm Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . 53
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
3.2 Genetic Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
3.2.1 Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
3.2.2 Representation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
3.2.3 Crossover and Mutation . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
3.2.4 Environmental Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
3.3 Real-Coded Genetic Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
3.3.1 Real-Valued Representation . . . . . . . . . . . . . . . . . . . . . . . . 60
3.3.2 Blended Crossover . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
3.3.3 Simulated Binary Crossover and Polynomial
Mutation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
3.4 Evolution Strategies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
3.4.1 (1 + 1)-ES . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
3.4.2 Evolution Strategies with One Global Step Size . . . . . . . 63
3.4.3 Evolution Strategies with Individual Step Sizes . . . . . . . 64
3.4.4 Reproduction and Environmental Selection . . . . . . . . . . . 64
3.4.5 Covariance Matrix Adaptation Evolution Strategy . . . . . 65
3.5 Genetic Programming . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
3.5.1 Tree-Based Genetic Programming . . . . . . . . . . . . . . . . . . . 67
3.5.2 Initialization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
3.5.3 Crossover and Mutation . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
3.6 Ant Colony Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
3.6.1 Overall Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
3.6.2 Extensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
3.7 Differential Evolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 73
3.7.1 Initialization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74
Contents xiii

3.7.2 Differential Mutation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 74

3.7.3 Differential Crossover . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
3.7.4 Environmental Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
3.8 Particle Swarm Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
3.8.1 Canonical Particle Swarm Optimization . . . . . . . . . . . . . . 76
3.8.2 Competitive Swarm Optimizer . . . . . . . . . . . . . . . . . . . . . . 78
3.8.3 Social Learning Particle Swarm Optimizer . . . . . . . . . . . 80
3.9 Memetic Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
3.9.1 Basic Concepts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
3.9.2 Lamarckian Versus Baldwinian Approaches . . . . . . . . . . 84
3.9.3 Multi-objective Memetic Algorithms . . . . . . . . . . . . . . . . 85
3.9.4 Baldwin Effect Versus Hiding Effect . . . . . . . . . . . . . . . . . 86
3.10 Estimation of Distribution Algorithms . . . . . . . . . . . . . . . . . . . . . . . 88
3.10.1 A Simple EDA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
3.10.2 EDAs for Discrete Optimization . . . . . . . . . . . . . . . . . . . . 90
3.10.3 EDAs for Continuous Optimization . . . . . . . . . . . . . . . . . . 91
3.10.4 Multi-objective EDAs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
3.11 Parameter Adaptation and Algorithm Selection . . . . . . . . . . . . . . . 94
3.11.1 Automated Parameter Tuning . . . . . . . . . . . . . . . . . . . . . . . 94
3.11.2 Hyper-heuristics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
3.11.3 Fitness Landscape Analysis . . . . . . . . . . . . . . . . . . . . . . . . 97
3.11.4 Automated Recommendation Systems . . . . . . . . . . . . . . . 97
3.12 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
4 Introduction to Machine Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
4.1 Machine Learning Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
4.1.1 Clustering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
4.1.2 Dimension Reduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
4.1.3 Regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
4.1.4 Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
4.2 Machine Learning Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
4.2.1 Polynomials . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
4.2.2 Multi-layer Perceptrons . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110
4.2.3 Radial-Basis-Function Networks . . . . . . . . . . . . . . . . . . . . 112
4.2.4 Support Vector Machines . . . . . . . . . . . . . . . . . . . . . . . . . . 113
4.2.5 Gaussian Processes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
4.2.6 Decision Trees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
4.2.7 Fuzzy Rule Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
4.2.8 Ensembles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 118
4.3 Learning Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
4.3.1 Supervised Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
4.3.2 Unsupervised Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
4.3.3 Reinforcement Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . 125
4.3.4 Advanced Learning Algorithms . . . . . . . . . . . . . . . . . . . . . 126
xiv Contents

4.4 Multi-objective Machine Learning . . . . . . . . . . . . . . . . . . . . . . . . . . 130

4.4.1 Single- and Multi-objective Learning . . . . . . . . . . . . . . . . 130
4.4.2 Multi-objective Clustering, Feature Selection
and Extraction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132
4.4.3 Multi-objective Ensemble Generation . . . . . . . . . . . . . . . . 135
4.5 Deep Learning Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135
4.5.1 Convolutional Neural Networks . . . . . . . . . . . . . . . . . . . . . 136
4.5.2 Long Short-Term Memory Networks . . . . . . . . . . . . . . . . 136
4.5.3 Autoassociative Neural Networks
and Autoencoder . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 137
4.5.4 Generative Adversarial Networks . . . . . . . . . . . . . . . . . . . 139
4.6 Synergies Between Evolution and Learning . . . . . . . . . . . . . . . . . . 140
4.6.1 Evolutionary Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141
4.6.2 Learning for Evolutionary Optimization . . . . . . . . . . . . . . 142
4.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 143
5 Data-Driven Surrogate-Assisted Evolutionary Optimization . . . . . . . 147
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147
5.2 Offline and Online Data-Driven Optimization . . . . . . . . . . . . . . . . 149
5.2.1 Offline Data-Driven Optimization . . . . . . . . . . . . . . . . . . . 149
5.2.2 Online Data-Driven Optimization . . . . . . . . . . . . . . . . . . . 150
5.3 Online Surrogate Management Methods . . . . . . . . . . . . . . . . . . . . . 151
5.3.1 Population-Based Model Management . . . . . . . . . . . . . . . 151
5.3.2 Generation-Based Model Management . . . . . . . . . . . . . . . 152
5.3.3 Individual-Based Model Management . . . . . . . . . . . . . . . 154
5.3.4 Trust Region Method for Memetic Algorithms . . . . . . . . 156
5.4 Bayesian Model Management . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 157
5.4.1 Acquisition Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 158
5.4.2 Evolutionary Bayesian Optimization . . . . . . . . . . . . . . . . . 159
5.4.3 Bayesian Evolutionary Optimization . . . . . . . . . . . . . . . . . 161
5.5 Bayesian Constrained Optimization . . . . . . . . . . . . . . . . . . . . . . . . . 162
5.5.1 Acquisition Function for Constrained
Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163
5.5.2 Two-Stage Acquisition Functions . . . . . . . . . . . . . . . . . . . 165
5.6 Surrogate-Assisted Robust Optimization . . . . . . . . . . . . . . . . . . . . . 166
5.6.1 Bi-objective Formulation of Robust Optimization . . . . . . 166
5.6.2 Surrogate Construction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167
5.7 Performance Indicators for Surrogates . . . . . . . . . . . . . . . . . . . . . . . 167
5.7.1 Accuracy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
5.7.2 Selection-based Performance Indicator . . . . . . . . . . . . . . . 168
5.7.3 Rank Correlation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 170
5.7.4 Fitness Correlation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 170
5.8 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 171
Contents xv

6 Multi-surrogate-Assisted Single-objective Optimization . . . . . . . . . . . 173

6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
6.2 Local and Global Surrogates Assisted Optimization . . . . . . . . . . . 174
6.2.1 Ensemble Surrogate Model . . . . . . . . . . . . . . . . . . . . . . . . . 175
6.2.2 Multi-surrogate for Single-objective Memetic
Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176
6.2.3 Multi-surrogate for Multi-objective Memetic
Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 176
6.2.4 Trust Region Method Assisted Local Search . . . . . . . . . . 178
6.2.5 Experimental Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 180
6.3 Two-Layer Surrogate-Assisted Particle Swarm
Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 181
6.3.1 Global Surrogate Model . . . . . . . . . . . . . . . . . . . . . . . . . . . 182
6.3.2 Local Surrogate Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . 183
6.3.3 Fitness Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184
6.3.4 Surrogate Management . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185
6.3.5 Experimental Results and Discussions . . . . . . . . . . . . . . . 185
6.4 Committee Surrogate Assisted Particle Swarm
Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 186
6.4.1 Committee of Surrogate Models . . . . . . . . . . . . . . . . . . . . 187
6.4.2 Infill Sampling Criteria . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188
6.4.3 Overall Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 188
6.4.4 Experimental Results on Benchmark Problems . . . . . . . . 189
6.5 Hierarchical Surrogate-Assisted Multi-scenario
Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 190
6.5.1 Multi-scenario Airfoil Optimization . . . . . . . . . . . . . . . . . 190
6.5.2 Hierarchical Surrogates for Multi-scenario
Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192
6.6 Adaptive Surrogate Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195
6.6.1 Basic Idea . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 195
6.6.2 Probabilistic Model for Surrogate Selection . . . . . . . . . . . 196
6.7 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199
7 Surrogate-Assisted Multi-objective Evolutionary
Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 201
7.1 Evolutionary Multi-objective Optimization . . . . . . . . . . . . . . . . . . . 201
7.1.1 Hypothesis and Methodologies . . . . . . . . . . . . . . . . . . . . . 201
7.1.2 Decomposition Approaches . . . . . . . . . . . . . . . . . . . . . . . . 203
7.1.3 Dominance Based Approaches . . . . . . . . . . . . . . . . . . . . . . 205
7.1.4 Performance Indicator Based Approaches . . . . . . . . . . . . 211
7.2 Gaussian Process Assisted Randomized Weighted
Aggregation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212
7.2.1 Challenges for Surrogate-Assisted Multi-objective
Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 212
xvi Contents

7.2.2 Efficient Global Optimization Algorithm . . . . . . . . . . . . . 214

7.2.3 Extension to Multi-objective Optimization . . . . . . . . . . . . 214
7.3 Gaussian Process Assisted Decomposition-Based
Multi-objective Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 216
7.3.1 MOEA/D . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 216
7.3.2 Main Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 218
7.3.3 Local Surrogate Models . . . . . . . . . . . . . . . . . . . . . . . . . . . 218
7.3.4 Surrogate Management . . . . . . . . . . . . . . . . . . . . . . . . . . . . 219
7.3.5 Discussions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221
7.4 High-Dimensional Multi-objective Bayesian Optimization . . . . . 221
7.4.1 Main Challenges . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 221
7.4.2 Heterogeneous Ensemble Construction . . . . . . . . . . . . . . . 222
7.4.3 Pareto Approach to Multi-objective Bayesian
Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223
7.4.4 Overall Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 224
7.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 228
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 228
8 Surrogate-Assisted Many-Objective Evolutionary
Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231
8.1 New Challenges in Many-Objective Optimization . . . . . . . . . . . . . 231
8.1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 231
8.1.2 Diversity Versus Preferences . . . . . . . . . . . . . . . . . . . . . . . 232
8.1.3 Search for Knee Solutions . . . . . . . . . . . . . . . . . . . . . . . . . 233
8.1.4 Solving Problems with Irregular Pareto Fronts . . . . . . . . 235
8.2 Evolutionary Many-Objective Optimization Algorithms . . . . . . . . 236
8.2.1 Reference Vector Guided Many-Objective
Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 236
8.2.2 A Knee-Driven Many-Objective Optimization
Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 240
8.2.3 A Two-Archive Algorithm for Many-Objective
Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 243
8.2.4 Corner Sort for Many-Objective Optimization . . . . . . . . 245
8.3 Gaussian Process Assisted Reference Vector Guided
Many-Objective Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 248
8.3.1 Surrogate Management . . . . . . . . . . . . . . . . . . . . . . . . . . . . 250
8.3.2 Archive Maintenance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 250
8.4 Classification Surrogate Assisted Many-Objective
Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 252
8.4.1 Main Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 252
8.4.2 Radial Projection Based Selection . . . . . . . . . . . . . . . . . . . 254
8.4.3 Reference Set Based Dominance Relationship
Prediction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 254
8.4.4 Surrogate Management . . . . . . . . . . . . . . . . . . . . . . . . . . . . 256
8.4.5 Surrogate-Assisted Environmental Selection . . . . . . . . . . 258
Contents xvii

8.5 Dropout Neural Network Assisted Many-Objective

Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 259
8.5.1 AR-MOEA . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 259
8.5.2 Efficient Deep Dropout Neural Networks . . . . . . . . . . . . . 262
8.5.3 Model Management . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 264
8.5.4 Overall Framework of EDN-ARMOEA . . . . . . . . . . . . . . 265
8.5.5 Operational Optimization in Crude Oil Distillation
Units . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 266
8.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 269
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 269
9 Knowledge Transfer in Data-Driven Evolutionary
Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 273
9.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 273
9.2 Co-Training for Surrogate-Assisted Interactive
Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 274
9.2.1 Overall Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 275
9.2.2 Surrogate for Interval Prediction . . . . . . . . . . . . . . . . . . . . 276
9.2.3 Fitness Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 278
9.2.4 An Improved CSSL . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 279
9.2.5 Surrogate Management . . . . . . . . . . . . . . . . . . . . . . . . . . . . 280
9.3 Semi-Supervised Learning Assisted Particle Swarm
Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 280
9.3.1 Algorithm Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 281
9.3.2 Social Learning Particle Swarm Optimization . . . . . . . . . 281
9.3.3 Surrogate Management Strategy . . . . . . . . . . . . . . . . . . . . 283
9.3.4 Selection of Unlabeled Data . . . . . . . . . . . . . . . . . . . . . . . . 284
9.3.5 Experimental Results and Discussions . . . . . . . . . . . . . . . 285
9.4 Knowledge Transfer between Problems in Multi-objective
Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 286
9.4.1 Domain Adaptation for Transfer Learning . . . . . . . . . . . . 286
9.4.2 Knowledge Transfer from Cheap to Expensive
Problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 290
9.4.3 CE-BDA for Data Augmentation . . . . . . . . . . . . . . . . . . . . 292
9.4.4 Evolutionary Multi-Objective Bayesian
Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 293
9.5 Knowledge Transfer between Objectives in Multi-objective
Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 294
9.5.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 294
9.5.2 Parameter Based Transfer Learning . . . . . . . . . . . . . . . . . . 294
9.5.3 Overall Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 295
9.6 Data-Driven Multi-fidelity Transfer Optimization . . . . . . . . . . . . . 296
9.6.1 Transfer Learning for Bi-Fidelity Optimization . . . . . . . . 297
9.6.2 Transfer Stacking . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 297
xviii Contents

9.6.3 Surrogate-Assisted Bi-Fidelity Evolutionary

Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 298
9.6.4 Experimental Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 300
9.7 Surrogate-Assisted Multitasking Multi-Scenario
Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 301
9.7.1 Multi-Scenario Minimax Optimization . . . . . . . . . . . . . . . 301
9.7.2 Surrogate-Assisted Minimax Multifactorial
Evolutionary Optimization . . . . . . . . . . . . . . . . . . . . . . . . . 302
9.7.3 Experiment Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 304
9.8 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 306
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 306
10 Surrogate-Assisted High-Dimensional Evolutionary
Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 309
10.1 Surrogate-Assisted Cooperative Optimization
for High-Dimensional Optimization . . . . . . . . . . . . . . . . . . . . . . . . . 309
10.1.1 RBF-Assisted SL-PSO . . . . . . . . . . . . . . . . . . . . . . . . . . . . 311
10.1.2 FES-Assisted PSO . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 312
10.1.3 Archive Update . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 316
10.1.4 Experimental Results and Discussions . . . . . . . . . . . . . . . 317
10.2 A Multi-objective Infill Criterion for High-Dimensional
Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 320
10.2.1 Main Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 321
10.2.2 Multiobjective Infill Criterion . . . . . . . . . . . . . . . . . . . . . . . 322
10.2.3 Experimental Results and Discussions . . . . . . . . . . . . . . . 325
10.3 Multi-surrogate Multi-tasking Optimization of Expensive
problems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 327
10.3.1 Multi-factorial Evolutionary Algorithms . . . . . . . . . . . . . 327
10.3.2 Main Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 328
10.3.3 Global and Local Surrogates . . . . . . . . . . . . . . . . . . . . . . . 329
10.3.4 Multi-tasking Optimization Based on Global
and Local Surrogates . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 331
10.3.5 Experimental Results and Discussions . . . . . . . . . . . . . . . 331
10.4 Surrogate-Assisted Large Optimization with Random
Feature Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 333
10.4.1 Main Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 335
10.4.2 Sub-problem Formation and Optimization . . . . . . . . . . . . 335
10.4.3 Global Best Position Update . . . . . . . . . . . . . . . . . . . . . . . . 337
10.4.4 Experimental Results and Discussions . . . . . . . . . . . . . . . 338
10.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 340
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 340
11 Offline Big or Small Data-Driven Optimization
and Applications . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 343
11.1 Adaptive Clustering for Offline Big-Data Driven
Optimization ... . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 343
Contents xix

11.1.1 Problem Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 343

11.1.2 Adaptive Clustering for Offline Data-Driven
Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 345
11.1.3 Empirical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 346
11.1.4 Discussions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 347
11.2 Small Data-Driven Multi-objective Magnesium Furnace
Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 348
11.2.1 Model Management Based on a Global Surrogate . . . . . 348
11.2.2 Empirical Verification on Benchmark Problems . . . . . . . 349
11.2.3 Optimization of Fused Magnesium Furnaces . . . . . . . . . . 350
11.3 Selective Ensemble for Offline Airfoil Optimization . . . . . . . . . . . 354
11.3.1 Problem Formulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 354
11.3.2 Selective Ensemble for Offline Data-Driven
Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 355
11.3.3 Comparative Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 356
11.4 Knowledge Transfer in Offline Data-Driven Beneficiation
Process ... . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 358
11.4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 358
11.4.2 Knowledge Transfer by Multi-surrogate
Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 359
11.4.3 Reference Vector Based Final Solution Selection . . . . . . 360
11.4.4 Optimization of Beneficiation Process . . . . . . . . . . . . . . . 361
11.5 Transfer Learning for Offline Data-Driven Dynamic
Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 364
11.5.1 Dynamic Data-Driven Optimization . . . . . . . . . . . . . . . . . 364
11.5.2 Data Stream Ensemble for Incremental Learning . . . . . . 365
11.5.3 Ensemble Based Transfer Optimization . . . . . . . . . . . . . . 367
11.5.4 Support Vector Domain Description for Final
Solution Selection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 367
11.5.5 Empirical Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 368
11.6 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 370
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 370
12 Surrogate-Assisted Evolutionary Neural Architecture Search . . . . . . 373
12.1 Challenges in Neural Architecture Search . . . . . . . . . . . . . . . . . . . . 373
12.1.1 Architecture Representation . . . . . . . . . . . . . . . . . . . . . . . . 374
12.1.2 Search Strategies . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 376
12.1.3 Performance Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . 377
12.2 Bayesian Optimization for Neural Architecture Search . . . . . . . . . 377
12.2.1 Architecture Encoding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 379
12.2.2 Kernel Functions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 379
12.2.3 Discussions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 380
12.3 Random Forest Assisted Neural Architecture Search . . . . . . . . . . 380
12.3.1 Block-Based Architecture Representation . . . . . . . . . . . . 381
12.3.2 Offline Data Generation . . . . . . . . . . . . . . . . . . . . . . . . . . . 382
xx Contents

12.3.3 Random Forest Construction . . . . . . . . . . . . . . . . . . . . . . . 383

12.3.4 Search Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 383
12.3.5 Experimental Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 384
12.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 385
References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 385

Index . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 389
Acronyms

ACO Ant colony optimization

AdaBoost Adaptive boosting
AIC Akaike information criterion
APD Angle penalized distance
AUE2 Accuracy updated ensemble algorithm
Bagging Bootstrap aggregating
BEO Bayesian evolutionary optimization
BO Bayesian optimization
BOA Bayesian optimization algorithm
CAL-SAPSO Committee-based surrogate-assisted particle swarm optimization
CART Classification and regression tree
CFD Computational fluid dynamic
CGA Compact genetic algorithm
CMA-ES Covariance matrix adaptation evolutionary strategies
C-MOGA Cellular multi-objective genetic algorithm
CNN Convolutional neural network
COBRA Constrained optimization by radial basis function interpolation
CPF Coverage over the Pareto front
CSEA Classification-based surrogate-assisted evolutionary algorithm
CSO Competitive swarm optimizer
CSSL Co-training semi-supervised learning
DAG Directed acyclic graph
DE Differential evolution
DSE Data stream ensemble
EA Evolutionary algorithm
EBO Evolutionary Bayesian optimization
ECT Electricity consumption for a ton of magnesia
EDA Estimation of distribution algorithm
EGO Efficient global optimization
EI Expected improvement
ES Evolution strategies
ESD Estimated standard deviation
xxi
xxii Acronyms

FCM Fuzzy clustering method

FDA Factorized distribution algorithm
FLOPs Floating point operations
GA Genetic algorithm
GAN Generative adversarial network
GD Generational distance
G-MFEA Generalized multifactorial evolutionary algorithm
GP Genetic programming
GP-MOEA Gaussian process-assisted multi-objective evolutionary algorithm
GS-MOMA Generalized surrogate-assisted multi-objective memetic algorithm
GS-SOMA Generalized surrogate-assisted single-objective memetic algorithm
HeE-MOEA Heterogeneous ensemble assisted multi-objective evolutionary
algorithm
HSA-MSES Hierarchical surrogate-assisted multi-scenario evolution strategy
HV Hypervolume
IBEA Indicator-based evolutionary algorithm
IGA Interactive genetic algorithm
IGD Inverted generational distance
KnEA Knee point-driven evolutionary algorithm
KNN K-nearest neighbours
LCB Lower confidence bound
LHS Latin hypercube sampling
LSTM Long short-term memory
MaOP Many-objective optimization problem
MBN Multidimensional Bayesian network
MFEA Multi-factorial evolutionary algorithm
MIC Multi-objective infill criterion
MLP Multi-layer perceptron
MOEA Multi-objective evolutionary algorithm
MOEA/D Multi-objective evolutionary algorithm based on decomposition
MOP Multi-objective optimization problem
MSE Mean squared error
MTO Multitasking optimization
NAS Neural architecture search
PBI Penalty boundary intersection
PBIL Population-based incremental learning
PCA Principal component analysis
PD Pure diversity
PoI Probability of improvement
PR Polynomial regression
PSO Particle swarm optimization
QBC Query by committee
RBF Radial basis function
RBFN Radial basis function network
RF Random forest
Acronyms xxiii

RL Reinforcement learning
RM-MEDA Regularity model-based multiobjective estimation of distribution
algorithm
RMSE Root mean square error
ROOT Robust optimization over time
SA-COSO Surrogate-assisted cooperative swarm optimization algorithm
SBX Simulated binary crossover
SGD Stochastic gradient descent
SL-PSO Social learning particle swarm optimization
SOM Self-organizing map
SOP Single-objective optimization problem
SP Spacing
SQP Sequential quadratic programming
SVDD Support vector domain description
SVM Support vector machine
TLSAPSO Two-layer surrogate-assisted particle swarm optimization
TSP Travelling salesman problem
UCB Upper confidence bound
UMDA Univariate marginal distribution algorithm
Symbols

R Real number space

N Integer space
 Feasible region
D Data set
x Decision variable or input variable (feature)
x Decision vector or input vector (solution)
y Label or output variable
y Label or output vector
n Number of decision variables or dimension of input
d Number of data
m Number of objective functions
f (x) Objective function
g(x) Inequality constraint function
h(x) Equality constraint function
fˆ(x) Approximated function
F(x) Function vector
ϕ Fidelity of approximate objective function
s Scenario
zideal or zå Ideal point of multi-objective optimization problem
znadir Nadir point of multi-objective optimization problem
zutopian Utopian point of multi-objective optimization problem
P Parent population
O Offspring population
N Population size
NF E Number of fitness evaluations
α Learning rate

xxv
Chapter 1
Introduction to Optimization

1.1 Definition of Optimization

Optimization problems can be found everywhere in engineering, economics, busi-

ness, and daily life. For instance, design of a turbine engine intends to minimize the
pressure loss, while design of airfoil or a race car usually aims to minimize the drag.
In process industries, maximization of product quality and quantity is desired, and
in automated trading or business investment, one is often interested in maximizing
the profit. Even in daily life, many decision-making processes such as purchasing
a personal computer or mobile phone requires to minimize the costs. In general,
optimization involves making the best use of given situation or resources.

1.1.1 Mathematical Formulation

Formally, optimization, also known as mathematical programming, is a mathematical

discipline that aims to find of the minimums and maximums of an objective function,
subject to certain constraints (Kochenderfer & Wheeler, 2019). An optimization
problem consists of three main components (Nocedal & Wright, 1999):
• Objective function: This defines the mathematical representation of the measure
of performance in terms of the decision variables. For example, a linear objective
function may look like:

minimize f (x1 , x2 ) = 4x1 − x2 ; (1.1)

• Decision variables: Each decision variable represents a quantifiable decision that

needs to be made. By changing the decision variables, the objective value will also
be changed. In Eq. (1.1), x1 and x2 are decision variables, and the value of the
objective function f (x1 , x2 ) will be changed by changing x1 or x2 .
© The Author(s), under exclusive license to Springer Nature Switzerland AG 2021 1
Y. Jin et al., Data-Driven Evolutionary Optimization,
Studies in Computational Intelligence 975,
https://doi.org/10.1007/978-3-030-74640-7_1
2 1 Introduction to Optimization

• Constraints: Constraints are mathematical equations or inequalities that indicate

restrictions or conditions on the values the decision variables can take. For example
in the above example, the optimal solution should satisfy the following conditions:

x1 + x2 = 10.0, (1.2)
x1 − x2 < 5.0 (1.3)
0 ≤ x1 , x2 < 10.0 (1.4)

Of the above two constraints, Eq. (1.2) is called an equality constraint, Eq. (1.3)
is called an inequality constraint, and Eq. 1.4 is known as boundary constraints that
defines the lower and upper bounds of the decision variables. For an optimization
problem, one combination of the decision variables is known as a solution. For
example, for the optimization problem in (1.1), (x1 = 5.0, x2 = 5.0) is solution of
the problem, and since this solution satisfies both constraints, it is called a feasible
solution. By contrast, (x1 = 1.0, x2 = 8.0) does not satisfy the equality condition, and
therefore it is known as a non-feasible solution. The task of the above optimization
problem is to find a feasible solution that minimizes the objective function (1.1).
Generally, a minimization problem can be formulated as

minimize f (x) (1.5)

subject to g j (x) < 0, j = 1, 2, . . . , J. (1.6)
h k (x) = 0, k = 1, 2, . . . , K . (1.7)
x ≤ x ≤ xU .
L
(1.8)

where f (x) is the objective function, x ∈ Rn is the decision vector, x L and xU are
the lower and upper bounds of the decision vector, n is the number of decision
variables, g j (x) are the inequality constraints, and h k (x) are the equality constraints,
J ,K are the number of inequality and equality constraints, respectively. If a solution
x satisfies all constraints, it is called a feasible solution, and the solutions that achieve
the minimum value are called the optimal solutions or minimums. For a differentiable
objective function where all its partial derivatives equal zero, the solution is called
a stationary point, which may be a local minimum, a local maximum, or a saddle
point, as illustrated in Fig. 1.1.
Problem formulation in solving real-world optimization problems is of extreme
importance but challenging. The main reasons include
• In optimization of complex systems, there are large numbers of decision variables,
objectives and constraints and it is unpractical to optimize the overall system at
once. For example, deigning a car may be divided into conceptual design, design
of components and design of parts. Note that the amount of time and cost allowed
for these different design stages are different.
• Even for a subsystem, it is not easy for the user to determine whether a particular
requirement should be considered as an objective or as a constraint. In addition,
some objectives or constraints are added only after some preliminary optimization
1.1 Definition of Optimization 3

Fig. 1.1 Illustration of

stationary points of a
one-dimensional function,
where solution ‘A’ is a saddle
point, ‘B’ and ‘D’ are local
minimums, and ‘C’ is a local
maximum

runs are performed. For example in designing a turbine engine, minimization of

the pressure loss is the primary concern. However, after a preliminary design is
optimized, it is found that the variance of flow rate along the outlet is too big,
which is detrimental. Thus, a second objective to minimize the variance, or an
additional constraint on the variance must be included in the problem formulation.

1.1.2 Convex Optimization

Definition 1.1 The optimization problem in Eq. (1.5) is called convex, if for all x1 ,
x2 ∈ Rn , and for all α + β = 1, α ≥ 0, β ≥ 0, the following condition holds:

f (αx1 + βx2 ) ≤ α f (x1 ) + β f (x2 ). (1.9)

If the optimization problem in Eq. (1.5) is convex, the objective function is a

convex function and the feasible solution set is a convex set. If a function f (x) is
convex, then − f (x) is concave.
An example of a convex function is given in Fig. 1.2a. Visually, the function is
convex if all points on the line connecting any two points (here x1 and x2 ) are above
the function.
A convex optimization problem may have none, one or multiple optimal solutions.
However, if the problem is strictly convex, then the problem has at most one optimal
solution. Note that a complex optimization problem can be solved in polynomial
time, whilst a non-convex optimization problem is generally NP-hard.
4 1 Introduction to Optimization

Fig. 1.2 a An example of a convex function. b An example of quasi-convex function, which is

convex for x ∈ Sα . However, f (x) is concave, when e.g., x1 ≤ x2

1.1.3 Quasi-convex Function

Definition 1.2 A function f (x) is said to be quasi-convex on S, where S is a convex

subset of R, the following condition satisfies:

f (λx1 + (1 − λ)x2 ) ≤ max{ f (x1 ), f (x2 )}. (1.10)

for all x1 , x2 ∈ S, and all λ = [0, 1].

Alternatively, function f (x) is quasi-convex on a sublevel set Sα = {x| f (x) ≤ α},

if Sα is convex.
An example of quasi-convex function is given in Fig. 1.2b, where f (x) is convex
on Sα . However, it is concave on the upper right part, as indicated by the line con-
necting f (x1 ) and f (x2 ), where all points on the line connecting f (x1 ) and f (x2 )
are below f (x).
Note that a strictly quasi-convex function is also known as unimodal. A unimodal
function has one local optimum, which is also the global optimum. By contrast,
functions having multiple local optimums are called multi-modal functions.

1.1.4 Global and Local Optima

Consider the following minimization problem:

minimize f (x) (1.11)

subject to x ∈ F (1.12)
1.1 Definition of Optimization 5

where f : Rn → R, n is the number of decision variables, F ⊂ Rn is the feasible

region. Then we have the following definitions.

Definition 1.3 A point x ∈ F is a global optimum of the optimization problem in

Eq. (1.11), if
f (x∗ ) ≤ f (x) (1.13)

∀x ∈ F.

Definition 1.4 A point x ∈ F is a local optimum of the optimization problem in Eq.

(1.11), if there exists ρ > 0 such that

f (x∗ ) ≤ f (x) (1.14)

∀x ∈ F and x − x∗ < ρ.

1.2 Types of Optimization Problems

Optimization problems can be categorized into various classes, typically according

to the nature of the decision variables, the number of objectives to be optimized,
whether constraints need to be considered, or whether uncertainty is present in the
optimization problems. Finally, there also optimization problems whose objective
functions are unknown or cannot be described by analytic mathematical equations.

1.2.1 Continuous Versus Discrete Optimization

A large number of real-world optimization problems continuous problems, in which

the decision variables are real numbers and they can be changed continuously in the
feasible regions. For example, geometry design of a turbine engine is a continuous
optimization problems, where the decision variables describing the shape of the
turbine are real numbers.
In contrast to continuous optimization problems, all or part of decision variables of
discrete optimization problems can take on discrete sets of values only, often integers.
For instance, number of people, number of cities or locations, and the sequence in
which a manufacturing process is organized. Discrete problems can further be divided
into mixed-integer problems, in which part of the decision variables are integers,
integer problems, and combinatorial problems. Combinatorial optimization can be
seen as a special class of integer optimization, where the task is to find an optimal
combination of a finite set of objects. Representative combinatorial problems include
the travelling salesman problem (TSP), the knapsack problem, and the minimum
spanning tree problem.
6 1 Introduction to Optimization

Fig. 1.3 An example of a continuous optimization problem. a Illustration of a jet engine turbine
blade. b A 3D B-spline representation of a blade using two 2D sections

Consider a simplified turbine engine optimization problem, where the task to

design the geometry of the turbine blades for energy efficiency. The problem can be
formulated as follows:

• Decision variables. Parameters defining the geometry of the turbine blade. Note
that many different methods can be used to represent a geometry, e.g., based on
parameterization or B-splines, as shown in Fig. 1.3. Here, the decision variables are
related to the blade geometry and therefore are continuous. So this is a continuous
optimization problem. There is typically a trade-off between completeness and
compactness in representing a design. A complete representation is able to repre-
sent all different structures, may may have a large number of decision variables.
By contrast, a compact representation has a small number of decision variables,
but may be limited in presenting complex structure. Other constraints may include
causality and locality (Jin & Sendhoff, 2009).
• Constraints. Here the constraints may include the mechanical constrains, among
others.
• Objectives. Usually, minimization of the pressure loss is the main target in design of
a turbine blade, which is closely related to energy efficiency. However, objectives
such as deviation from a desired inflow angle, or the variance of the flow rate at the
out let, may also be included in optimization. Note that the pressure loss cannot
be calculated using an analytic mathematical function and can be evaluated using
numerical simulations or wind tunnel experiments only.

The TSP is a classical combinatorial optimization problem which aims to find the
shortest path connecting cities visited by a travelling salesman on his sales route.
Each city can be visited only once. The TSP can be formulated as follows.
• Decision variables. Here the connecting order of the cities to be visited is the
decision variable. For example, in the 10-city example shown Fig. 1.4, one solution
is A → B → C → D → E → F → G → H → I → J → A.
1.2 Types of Optimization Problems 7

Fig. 1.4 An example of the

travelling salesman problem
with 10 cities showing a path
starting from city A and back
to A: A → B → C → D →
E→F→G→H →
I →J→A

• Constraints. The constrains are each city can be visited only once and the salesman
must be back to the starting city A is he starts from there.
• Objectives. The objective is to minimize the length of the total travelling path.

1.2.2 Unconstrained Versus Constrained Optimization

In unconstrained optimization problems, no equality or inequality constrains need

to be taken into account. By contrast, constrained optimization must satisfy equality
and/or inequality constraints, making the problem-solving harder. A special case of
constrained problems are box-constrained problems, in which only constraints on the
range of the decision variables are limited. Box-constrained optimization problems
reduce to unconstrained optimization when the bounds are infinite.

1.2.3 Single Versus Multi-objective Optimization

The problem in Eq. (1.5) has one objective function, therefore, such optimization
problems are known as single-objective optimization problems (SOPs). In the real
world, however, most problems have more than one objective to optimize, and these
objectives are usually conflicting with each other. These problems are called multi-
objective optimization problems (MOPs). In the evolutionary computation commu-
8 1 Introduction to Optimization

nity, MOPs with more than three objectives are termed many-objective optimiza-
tion problems (MaOPs), mainly because the hardness to solve MaOPs will become
dramatically more challenging for algorithms designed for solving two- or three-
objective problems.
There are also cases where the task of optimization is to find a feasible solution for
a highly constrained optimization problem, which is called constraint satisfaction.

1.2.4 Deterministic Versus Stochastic Optimization

Usually it is assumed that a given optimization is time invariant and there is no uncer-
tainty in the objective or constraint functions, nor in the decision variables. These
optimization problems are called deterministic optimization problems. In contrast to
deterministic problems, stochastic problems are subject to uncertainty in the objec-
tive and/or constraint functions, and in decision variables. The uncertainty may come
from different sources, such as noise in sensors, randomness in performing numerical
simulations, and changes in the operating condition. In these cases, one is interested
in finding a solution that is less sensitive to the uncertainty, which is an important
research topic of robust optimization and dynamic optimization to be elaborated in
Sect. 1.4.
It should be noted that stochastic optimization may also refer to research method-
ologies in which the decision variables are treated as random variables or the search
allow random changes. Here by an stochastic problem, we mean an optimization
problem that is subject to various types of uncertainty.

1.2.5 Black-Box and Data-Driven Optimization

In the discussions so far, we assume that the objective of an optimization problem can
be described analytically by a mathematical expression. However, this is not always
possible for solving real-world problems, where the objective function is unknown
or cannot be written in an analytic form. These optimization problems are sometimes
called black-box optimization.
Typically, the objective value of a black-box optimization problem is evaluated
using computer simulations. There are also cases when the objective is evaluated by
performing physical or chemical experiments, although the system to be optimized is
not necessarily a black-box. For example, the aerodynamic performance of a turbine
engine or a vehicle is well well understood scientifically, although the aerodynamic
performance of such systems can be evaluated only by solving a large number of
differential equations. In addition, in the era of data science and big data, many
optimization problems are solved based on collected data, which may be incomplete,
noisy, and heterogeneous. Therefore, these problems are better called data-driven
1.2 Types of Optimization Problems 9

optimization problems (Jin et al., 2018), where optimization relies on data collected
from computer simulations, physical or chemical experiments, and data from daily
collected life.

1.3 Multi-objective Optimization

1.3.1 Mathematical Formulation

In real-life, almost all optimization problems have multiple objectives to be max-

imized or minimized. For example, when one wants to buy a mobile phone, one
needs to consider the price that can be afforded, and the performance that is desired.
For performance, there are again multiple aspects to be taken into account, e.g.,
the specifications of the camera, the easiness to use and many others. It is practi-
cally impossible to simultaneously optimize all objectives, because there is usually
a trade-off between different objectives. For instance, as illustrated in Fig. 1.5, there
is a compromise, also known as trade-off, between price minimization and the per-
formance maximization. In other words, the better the performance, the higher the
price will be. Unlike in single-objective optimization where one optimal solution can
be obtained, there is always a set of trade-off solutions. Finally, one solution may
be selected based on the user’s preference. For example, if the user is willing to pay
more for a better mobile phone, he or she will select solution C. If the user wants to
have a good price/performance ratio, the user may select solution B. If the user put
more importance on the minimization of the cost, he or she will select solution A.
Mathematically, a multi-objective optimization problem can be formulated as
follows (Miettinen, 1999):

Fig. 1.5 Purchase of a

mobile phone is a
bi-objective optimization
problem that maximizes the
performance and minimizes
the price. A set of optimal
trade-off solutions will be
obtained instead of a single
ideal solution
10 1 Introduction to Optimization

minimize f i (x) i = 1, 2, . . . , m. (1.15)

subject to g j (x) < 0, j = 1, 2, . . . , J. (1.16)
h k (x) = 0, k = 1, 2, . . . , K . (1.17)
xlL ≤ xl ≤ xlU , l = 1, 2, . . . , n. (1.18)

where m > 1 is the number of objectives. If m = 1, it reduces to the single-objective

optimization problem described in Eq. (1.5).

1.3.2 Pareto Optimality

The comparison between solutions is different from that in single-objective opti-

mization, as more than one objective needs to be taken into account. For this reason,
a concept of Pareto dominance is introduced and is the most important definition in
multi-objective optimization.
Definition 1.5 For the minimization problem in Eq. (1.15), solution x1 Pareto dom-
inates, denoted by x1 ≺ x2 , solution x2 if

∀i ∈ {1, 2, . . . , m}, f i (x1 ) ≤ f i (x2 ), and (1.19)

∃k ∈ {1, 2, . . . , m}, f k (x1 ) < f k (x2 ). (1.20)

For example in Fig. 1.6, solution A dominates solutions C and D, while solution
B is not dominated by solution A. On the other hand, solution B does not dominate
A either. Therefore, solutions A and B are non-dominated, and similarly, solutions
B, C and D, do not dominate one another. If solutions A and B are not dominated
by any feasible solutions, they are Pareto optimal solutions.
Definition 1.6 A solution x∗ is called Pareto optimal, if there does not exist any
feasible solution x ∈ Rn such that x ≺ x∗ .

Fig. 1.6 Solution

comparison in
multi-objective optimization,
where solution A dominates
C and D, but does not
dominate solution B
1.3 Multi-objective Optimization 11

The image of all Pareto optimal solutions in the objective space is called the
Pareto front or Pareto frontier, and the set of all solutions in the decision space is the
Pareto optimal set consisting of two connected piece-wise linear curves, as shown in
Fig. 1.7. In the figure, a few particular points are often used in optimization, which
are the ideal point, the nadir point and the knee point.
Definition 1.7 The ideal point, denoted by zideal , is constructed from the best objec-
tive values of the Pareto set.
Definition 1.8 The nadir point, denoted by znadir , is constructed from the worst
objective values over the Pareto optimal set.
In practice, a utopian point is often defined as follows:
utopian
zi = z iideal − , for all i = 1, 2, . . . , m (1.21)

where  is a small positive constant.

Definition 1.9 A knee point is defined to be the one having the maximum distance
from the convex hull of individual minimums to the hyperplane constructed by the
extreme points.
In Fig. 1.7, solutions A and B are called the boundary points of the Pareto front,
the line connecting these two solutions is the hyperplane, and the intersections of
the hyperplane with the axes, here points D and E are called the extreme solutions.
Solution C has the maximum distance to the hyperplane, and therefore is the knee
point. Knee points of a Pareto front are of interest since they need a large compromise
in at least one objective to gain a small improvement in other objectives.
Note that there are also other definitions for knee points (Deb & Gupta, 2011). In
addition to the above distance based definition, there are also angle-based approaches
measuring the angle between a solution and its two neighbors. Although angle-based
approaches are straightforward, they are only applicable to bi-objective optimization
problems. By contrast, distance-based approaches can handle problems with more
than two objectives. Nevertheless, specifying knee points is notoriously difficult in
high-dimensional objective spaces, because a Pareto front may have a large number
of knee points.
In addition to the standard Pareto dominance relation, other dominance relation-
ship have also been suggested by modifying the Pareto dominance. In the following,
we introduce the -dominance (Laumanns et al., 2002) and α-dominance (Ikeda
et al., 2001) that are often used in evolutionary multi-objective optimization.
Definition 1.10 A solution x1 -dominates solution x2 for some  > 0, if

∀i ∈ {1, 2, . . . , m}, (1 − ) f (x1 ) ≤ f (x2 ). (1.22)

Definition 1.11 Given two solutions x1 , x2 ∈ Rn and x1 = x2 , solution

x1 α-dominates x2 , denoted by x1 ≺α x2 , if
12 1 Introduction to Optimization

Fig. 1.7 Illustration of the Pareto set (left panel) and Pareto front (right panel), together with the
ideal point, nadir point, knee point (C), boundary points (A and B) and extreme points (C and D)

Fig. 1.8 Illustration of dominance relations. a Pareto dominance, b -dominance, and c α-

dominance

∀i ∈ {1, 2, . . . , m}, gi (x1 , x2 ) ≤ 0, and (1.23)

∃ j ∈ {1, 2, . . . , m}, g j (x1 , x2 ) < 0, (1.24)

where

m
gi (x1 , x2 ) = f i (x1 ) − f i (x2 ) + αi j ( f j (x1 ) − f j (x2 )), (1.25)
j =i

and ai j is a parameter defining the trade-off rate between the i-th and j-th objectives.

From the above definitions, we can see that both -dominance and α-dominance
strengthen the Pareto dominance relation, enabling a solution to dominate more
solutions.
An illustration of Pareto dominance, -dominance and α-dominance relations is
given in Fig. 1.8. In Fig. 1.8a, the region solution A Pareto dominates is shaded and
therefore, solution B is not dominated by A. Figure 1.8b shows the region solution
A -dominates, which is larger than that is dominated by the Pareto dominance.
1.3 Multi-objective Optimization 13

Fig. 1.9 Different types of Pareto optimal fronts

Nevertheless, solution A still cannot dominate B. The region solution A α-dominates

is shown in Fig. 1.8c, where we can see that now solution B is dominated solution A.
From these examples, we can see that both -dominance and α-dominance relations
can dominate more the standard Pareto dominance, therefore, they can accelerate
the convergence of an optimization algorithm that uses the dominance relation for
comparing solutions.
The Pareto front of a multi-objective optimization problem may have very differ-
ent natures, which heavily influences the hardness of problem-solving. Figure 1.9
illustrates a few typical types of Pareto fronts. Note that the definition of convexity
of Pareto fronts is similar to the convexity of functions. For example, in Fig. 1.9a,
the Pareto front is convex, because all points on the line connecting any two points
on the Pareto front are dominated by the Pareto optimal solutions between these two
points. By contrast, all solutions connecting two points on a concave Pareto front
are infeasible solutions. In Fig. 1.9e the Pareto front is called degenerate, as for a
three-objective optimization problem, the dimension of a Pareto front is usually a
two-dimensional surface. Here, it degenerates to a one-dimensional curve.
14 1 Introduction to Optimization

1.3.3 Preference Modeling

The solution to a multi-objective optimization problem is a Pareto optimal set, which

may consist of an infinite number of solutions. In practice, however, only a small
subset of the Pareto optimal solutions will be adopted for implementation, for which
user preferences are needed. Two important questions must be answered to achieve
this, namely, how the user’s preferences should be properly represented and when
the preferences should be embedded in an optimization algorithm to approximate
the preferred solutions (Wang et al., 2017b).
Generally speaking, user preferences can be represented using the following meth-
ods.
• Goals. The most straightforward way to articulate user preferences is to provide
goal information (Gembicki, 1974), typically in terms of reference solutions or
reference points. In order words, the user can describe their preferences by provid-
ing one or multiple desired solutions. For example in optimizing a diesel engine,
the user may specify the desired fuel consumption and emission, which might be
a certain percentages better than the baseline design.
• Weights. For a completely new problem, it can be hard for the user to provide
desired solutions (goals) for different objectives. However, the user may be able
to assess the importance of different objectives. Since there is always a trade-off
between different objectives, it might be easier for the user to assign different levels
of importance to different objectives by using a weight vector w = {w1 , . . . wm },
where m is the number of objectives. Given the weights, a multiple objective
optimization problem can then be scalarized into a single-objective optimization
problem. The most widely used scalarizing is the linear scalarization, also known
as the weighted aggregation method (Miettinen, 1999):


m
g(x) = wi f i (x) (1.26)
i=1

where wi > 0, i = 1, 2, . . . , m and optimal solutions of the above single-objective

optimization problem are Pareto optimal solutions of the multi-objective optimiza-
tion problem in (1.15).
The linear scalarizing method is relatively easy to implement, however, has sev-
eral issues. The biggest problem is that one is not able to achieve Pareto optimal
solutions on the concave part of the Pareto front by solving the single-objective
problem in (1.26), no matter how the weights are tuned.
Another widely used scalarizing method is called the Chebyshev scalarizing func-
tion, which is expressed as follows (Miettinen, 1999):

g(x) = max {wi f i (x)}, (1.27)

1≤i≤m

where f i (x) is the i-th objective and wi is the i-th weight.

1.3 Multi-objective Optimization 15

Similar to goals, however, it is hard for the user to to specify accurate weights
without a good understanding the problem. It is particularly tricky when the Pareto
front is non-convex, or non-uniform, discrete, or degenerate.
• Reference vectors. Reference vectors (Wang et al., 2017b) are similar to goals
and weights in that they are meant to provide an expectation of the objectives.
Different to the goals (reference points), reference vectors provide information
about the directions in the objective space in which solutions are desirable. Sim-
ilar to weights, reference vectors can also be used to convert a multi-objective
optimization problem into single-objective optimization problems.
• Preference relation. Another natural way of preferences over different objectives
is to describe the relative importance between pairs of objectives using human lin-
guistic variables. For example, one may indicate that objective 1 is more important
than objective 2, or objective 1 and objective 2 are equally important.
In optimization, a preferred order of the objectives can be converted into weights
or weight intervals. However, one main disadvantage is that the preference relation
cannot handle non-transitivity.
• Utility functions. Utility functions can be used to represent user, where the prefer-
ence information is implicitly involved in the objective function to rank solutions.
Unlike preference relations, the utility function ranks solutions rather than the
objectives. For example, given N solutions x1 , x1 , . . ., x N , the user is required to
give his or her preferences over the solutions by ranking them in a order. Then, an
imprecisely specified multi-attribute value theory formulation is employed to infer
the relative importance of the objectives. However, utility functions are based on
a strong assumption that all attributes of the preferences are independent, thereby
being unable to handle non-transitivity.
• Outranking. Neither preference relations based on the importance of objectives
nor the utility functions based on solutions are able to handle non-transitivity. An
alternative is so-called outranking that allows for non-transitivity. To determine an
outranking, a preference and indifference thresholds for each objective are given
by a preference ranking organization method for enrichment evaluations (Brans
et al., 1986), according to which every two solutions are compared. Consequently,
a preference ranking is obtained, which can be used for search of preferred solu-
tions. However, the outranking based methods require a large number of parameter
settings, which is non-trivial for the user for problems having a large number of
objectives (Brans et al., 1986).
• Implicit preferences. Articulation of preferences is always challenging for the
user when there is little knowledge about the problem to be solved. In this case,
knee solutions, around which a small improvement of any objective causes a large
degradation of others, are always of interest to to the user. In addition to knee points,
extreme points or the nadir point can work as a special form of preferences. With
the help of extreme points or the nadir point, the user can acquire knowledge about
the range of the Pareto front so as to describe their preferences more accurately.
16 1 Introduction to Optimization

1.3.4 Preference Articulation

Once the user preferences are properly modeled, they must be included in the opti-
mization process to obtain the preferred solutions, although this is not straightfor-
ward. In the following, we elaborate various preference articulation methods widely
used in multi-objective optimization.

• No-preference. The simplest case is that no particular preferences are included in

the optimization and all objectives are treated equally important. For example, the
following scalarizing function can be used:

minimize f (x) − zideal , (1.28)

subject to x ∈ Rn . (1.29)

where · is usually L 1 , L 2 or L ∞ . Usually, the objectives need to be normalized

into a uniform scale so that no strong bias is included in the optimization.
• A priori approaches. A priori approaches to multi-objective optimization embed
the preferences in the objective functions before the optimization starts so that a
small subset of preferred Pareto optimal solutions are obtained. Most preference
models, such as goals (reference points), weights, and reference vectors can be
adopted to guide the search. In addition to the above-mentioned linear and Cheby-
shev scalarization methods that use weights for modeling preferences, we also
have the augmented weighted Chebyshev scalarizing method:

utopian 
m
utopian
min max wi [| f i (x) − z i |]+ ρ | f i (x) − z i |, (1.30)
i=1,2,...,m i=1
subject to: x ∈ Rn . (1.31)

where ρ > 0 is a small constant.

Alternatively, if the user can provide a reference point z̄, then the achievement
scalarizing function can be used:
 
f i (x) − z̄ i 
m
f i (x)
min max +ρ utopian , (1.32)
utopian z inadir −z i
i=1,2,...,m z inadir − zi i=1

subject to: x ∈ Rn . (1.33)

In principle, most classical optimization methods that need to convert a multi-

objective optimization problem into a single-objective problem using a scalarizing
function belong to a priori methods. A priori approaches suffer from two main
weaknesses. On the one hand, it is non-trivial for the user to provide informa-
tive and reasonable preferences prior to the optimization. On the other hand, an
optimization algorithm may not be able to find the preferred solutions even if the
user’s preferences are properly modelled.
1.3 Multi-objective Optimization 17

• A posteriori approaches. In the A posteriori approaches, the user is allowed to

articulate his or her preferences after a set of non-dominated solutions has been
obtained. To this end, the algorithms shall be able to approximate a representative
subset of the whole Pareto optimal front so that the user is able to select the preferred
ones from the obtained solutions. Since mathematical programming methods can
get one Pareto optimal solution in one run, they need to be run for a sufficient large
number of times to get a solution set. In contrast to mathematical programming
methods, evolutionary algorithms and other population based metaheuristics are
able to obtain a set of Pareto optimal or non-dominated solutions in one single
run. However, since the theoretical Pareto front is unknown beforehand, these
algorithms aim to promote the diversity of the obtained solutions in an effort to
get a representative subset of Pareto optimal solutions.
It should be pointed out that the assumption that a population based optimization
algorithm is able to obtain a representative subset of Pareto optimal solutions may
practically fail to hold given limited computational resources, in particular when
the Pareto front is discrete, degenerate, or when the number of objectives becomes
large.
• Interactive approaches. Considering the difficulty in precisely modeling the user
preferences in the a priori approach, and the challenges in acquiring a represen-
tative subset of Pareto optimal solutions, interactive approaches, also known as
progressive approaches, aim to progressively embed the user’s preferences in the
course of the optimization so that the user is able to refine or even change his or
her preferences according to the solutions obtained.
One main challenge of the interactive approaches is that the human user needs to be
directly involved in the optimization, which may become infeasible due to human
fatigue. One possible solution to address this issue is to partly automatize the
preference refinement or modification process using machine learning techniques.

1.4 Handling Uncertainty in Optimization

Most real-world optimization problems are subject to different degree of determin-

istic and probabilistic uncertainty (Jin & Branke, 2005; Liu & Jin, 2021). One most
widely seen uncertainty is introduced in evaluating the objective functions introduced
by various sources, such as measurement noises or numerical errors. In the real-world,
optimization is often subject to more systematic uncertainty. For example, design of
a turbine engine must consider its performance in various operating conditions when
the velocity, weight and altitude if the aircraft changes. These time-varying factors
are not decision variables, but will influence the objective functions. On the other
hand, there might be small errors introduced in manufacturing the turbine blades or
the blades may slowly worn out in the lifetime. Ideally, the performance of the tur-
bine blade should be relatively insensitive to such changes in the operating condition
(called environmental parameters hereafter) or in the decision variables. Without loss
18 1 Introduction to Optimization

of generality, the objective function of a non-constrained single-objective optimiza-

tion problem in the presence of uncertainty can be described as follows:

min f (x + x, a + a) + z, (1.34)

where x, a are perturbations in the decision variables and environmental param-
eters, respectively, and z ∼ N (0, σ 2 ) is additive noise, σ 2 is the variance of the
noise.
Different methods have been proposed to handle different uncertainty in opti-
mization. For additive noise in fitness evaluations, the typical method is to evaluate
the fitness multiple times to reduce the influence of the noise (Rakshit et al., 2017).
For addressing the non-additive perturbations in the decision variables and envi-
ronmental parameters, one approach is to find a solution that is robust to changes
in environmental parameters or in the decision variables, which is typically known
as robust optimization. However, if the operating conditions change significantly,
a more realistic and effective approach will be multi-scenario optimization. If the
changes cannot be captured by a probability distribution, for example, there are con-
tinuous or periodic changes in the environment or decision variables, then dynamic
optimization may be more effective. Finally, robust optimization over time, which
aims to make a best compromise between performance and robustness, as well as
the cost that may occur in switching the designs.

1.4.1 Noise in Evaluations

The most straightforward method to cancel out the addition noise in fitness evalua-
tions is to do averaging either over time or over space, meaning sampling multiple
times of the same solution, or sampling multiple solutions around the solution, respec-
tively, and then use the average of the sampled objective values as the final objective
value (Liu et al., 2014).
Since evolutionary algorithms are population based search methods, other
approaches to noise reduction have also been proposed, including using a large
population size together with proportional selection, introducing recombination, and
introducing a threshold in selection. Since fitness evaluations can be computationally
intensive, noise reducing based on re-sampling or using a large population can be
impractical. Thus, efficiently re-sampling during the optimization becomes impor-
tant. For example, adapting the sample size is an effective approach. Typically, one
can use a smaller sample size in the early search stage and a larger sample size in
the later stage; alternatively, a large sample size can be used for a more promising
solution, and a small sample size is used for a poor solution. Apart from adaptive
sampling, use of local regression models to estimate the true fitness value is also
possible.
1.4 Handling Uncertainty in Optimization 19

Fig. 1.10 Illustration of robust solutions against changes in the decision variable (left panel) and
the environmental parameter (right panel)

1.4.2 Robust Optimization

The robustness of the solutions against perturbations in the decision variables and
environmental parameters is a basic requirement in practice. Figure 1.10 illustrates
two different situations of robustness. In the left panel, A and B are two local optima of
the objective function. Without considering the robustness of the solutions, solution
A has better performance than solution B, if we are considering a minimization
problem. However, when there is a perturbation x in the decision variable, f (x A )
will be much worse than f (x B ). Therefore, we consider solution B is more robust
to solution A. By contrast, in the right panel, solution A has the best performance at
the normal operating condition when a = a ∗ . However, when a becomes larger or
smaller, the performance worsens very quickly. In contrast solution A, solution B
has worse performance than solution A at the normal operating condition, however,
its performance degrades more slowly. As a result, solution B performs better than
A over a ∈ R2 except for a ∈ R1 . Thus, solution B is more robust than solution A
against changes in the environmental parameter a. Before we discuss methods for
obtaining robust optimal solutions, we first introduce a few widely used definitions
for robustness in the context of optimization (Beyer & Sendhoff, 2007; Jin & Branke,
2005).

• Robust counterpart approach. For the optimization problem in the presence of

uncertainty in the decision variables, the robust counterpart function F(x, ) is
defined by
F(x, ) = supξ ∈R(x,) f (ξ ), (1.35)

where R(x, ) is a neighborhood of x of size . Typically, the neighborhood of x

of size  is defined by x ± . Therefore, this definition is the worst case scenario,
which is also known as a minimax optimization problem.
20 1 Introduction to Optimization

If we consider the perturbations to the environmental parameters, then the robust

counterpart function is given by

FW (x, ) = supa∈A() f (x, a), (1.36)

where A() is a neighborhood of a of size .

• Expected objective function. The expected objective function considering pertur-
bations in both decision variables and environmental parameters, we have

F(x) = f (x + δ, a) p(δ, a)dδda, (1.37)

If we consider the perturbations in the decision variables only, the expected objec-
tive function will be:

FE (x) = f (x + δ, a) p(δ)dδ. (1.38)

• Dispersion based robustness measure. The robustness can also be defined using
the dispersion of the objective function in the presence of uncertainty: e.g.

FD (x) = ( f (x + δ) − f (x))2 pδdδ, (1.39)

Methods for search of robust optimal solutions are based on the above robustness
definitions and their variants. That is, if one is interested in finding the robust optimal
solution, then, one of the above robustness measure can be used to replace the original
objective function. Since estimation of the robustness, no matter which of the above
definitions is used, requires to evaluate the objective value multiple times for one
solution, several ideas for reducing the computational costs have been proposed by
taking advantage of the population based search algorithm. For example, for the
expected objective function, the explicit averaging approach can calculate the mean
objective value to approximate the expected objective:

1 
N
f¯(x) = f (x + δ), (1.40)
N i=1

where δ ∼ N (0, σ 2 ) is a Gaussian noise, N is the sample size. Thus, N − 1 addi-

tional function evaluations are needed for evaluating each solution. To reduce the
computational cost, an implicit averaging method has been suggested, in which each
solution in the currently population is evaluated only once:

f¯(x) = f (x + δ), (1.41)

1.4 Handling Uncertainty in Optimization 21

where δ ∼ N (0, σ 2 ) is a Gaussian noise. Actually, the implicit approach is a

special case of the explicit averaging approach, where the sample size equals one.
This implicit works better when the population size of a population based search
algorithm is larger.
In the expected objective approach, it is assumed that the uncertainty is a Gaussian
noise, which requires to specify the level of the noise (variance). This is not easy,
and a wrong noise level used for calculating the robustness may either find a too
conservative solution, or a solution that is not robust enough. Thus, the worst case
scenario removes the assumption of Gaussian noise, but still requires the definition
of the neighborhood size and a large number of extra function evaluations to find out
the worst objective in the neighborhood. A variant of the worst scenario robustness,
called inverse robustness that aims to avoid the specification of the neighborhood
size, has been suggested. In the inverse robust optimization, the user can specify
the maximum tolerable performance degradation, based on which the maximum
allowed change in the decision variable will be found out. In practice, specification
the maximum tolerable performance degradation is more realistic than specifying
the noise level.
As can be seen in Fig. 1.10, there is usually a trade-off between the normal objec-
tive value and the degree of robustness (Jin & Sendhoff, 2003). Thus, robustness opti-
mization has often been addressed using the multi-objective optimization approach.
For example, one can minimize the expected objective function and minimize the
variance:

1 
N
FE (x) = f (xi ) (1.42)
N i=1

1 
N
FD (x) = [ f (xi ) − FE (x)]2 . (1.43)
N i=1

where N is the sample size.

Other combinations can also be used as the two objectives, such as the normal
objective function and the expected objective function, or the normal objective func-
tion together with the variance.
Once a set of non-dominated robust solutions are obtained, the user eventually
needs to select one solution a posteriori once any knowledge about the noise or the
tolerable performance degradation is available.

1.4.3 Multi-scenario Optimization

In practice, the performance of a design in several operating conditions must be

considered. For example for aircraft or turbine design, one must consider at least
three scenarios, take-off, cruise, and landing. For designing a vehicle, the dynamics
22 1 Introduction to Optimization

Fig. 1.11 The wave drag coefficient distribution of three RAE5225 airfoil designs obtained by
single- and multi-scenario optimization methods

of the vehicle in multiple driving scenarios must be considered, such as driving in a

city, on highway, on a curve, or passing a bridge. In a more extreme case, an infinite
number of scenarios must be considered, for example in designing a high-lift airfoil
system, where one aims to maximize the lift and minimize the drag for a given range
of flying velocity. This is also known as full scenario optimization (Wang et al.,
2018a). Figure 1.11 shows the drag coefficient distribution over a range of velocity
(from 0.5 mach number to 0.75 mach number) of three optimized designs of airfoil
RAE5255. From these figures, we can see different design philosophy may lead to
very different performances.
Several approaches can be adopted to deal with multi-scenario optimization. The
simplest approach is to aggregate the objectives in all considered scenarios. Alter-
natively, the objectives in each scenario are considered as separate objectives, and
consequently, the number objectives can become very large when multi-scenarios
are taken into account. In addition, the final decision making will also be difficult
without any preference over the scenarios.
Robustness optimization can also be used to tackle multi-scenario optimization,
where the averaging performance in multiple scenarios is optimized. However, this
will not lead to good performance if the number of scenarios is large.

1.4.4 Dynamic Optimization

Handling uncertainty by finding an optimal solution that is relatively insensitive to

uncertainty is effective if the level of the uncertainty is limited and can be described
by a certain type of Gaussian or uniform noise. However, this approach will fail if
the level of uncertainty is deterministic, continuous and large. In this case, the target
becomes to dynamically track the changing optimum or the changing Pareto optimal
solutions, which is known as dynamic optimization.
Generally, a dynamic optimization problem can be described as follows:
1.4 Handling Uncertainty in Optimization 23

min f (x, a(t)), (1.44)

where a(t) is a set of time-varying environmental parameters. In other words,

dynamic optimization is mainly concerned with uncertainty in the environmental
parameters. Although a(t) may change continuously over time, one basic hypothesis
in dynamic optimization is that the optimization problem can be seen as a sequence
of stationary problems (Branke, 2002; Jin & Branke, 2005):

< f (x, a1 ), f (x, a2 ), . . . , f (x, a L ) > (1.45)

where f (x, ai ), i = 1, 2, . . . , L is a time-invariant (stationary) function in the i-th

environment, L is the total number of environments. The total time interval for the
l environments is assumed to be [0, T ]. If the frequency of environmental change is
τ , then each L = T /τ .
The most straightforward strategy of solving the above dynamic optimization
problem is to restart the optimization once an environmental change is detected.
However, the restarting strategy is uninteresting and inefficient. Thus, another basic
assumption in dynamic optimization is that the optimization problem f (x, ai ) in the
i-th environment is more or less related to the one in the i − 1-th environment or those
in more than one previous environment. Thus, knowledge about the problems in the
previous environment is helpful in solving the problem in the current environment
more efficiently.
Many evolutionary dynamic optimization algorithms for transferring knowledge
in previous environments into the current environment have been proposed, which
can be divided into the following categories (Branke, 2002; Nguyen et al., 2012):
• Maintain or introduce diversity in the population. In solving stationary single-
objective optimization problems, an optimizer, e.g., an evolutionary algorithm,
will typically converge to one point once it finds an optimum. Loss of diversity
reduces the search capability of an optimizer if there is a change in the location
of the optimum. Thus, a straightforward strategy is to maintain a certain degree
of diversity in the population so that once an environmental change occurs, the
population can react quickly to find the new optimum. Alternatively, diversity is
introduced into the population once a change is detected.
Maintaining diversity in the genotype while achieving fast convergence is an inter-
esting phenomenon that can also be found in biology. It has been revealed that this
capability in biological systems can be partly attributed to phenotypic plasticity in
the lifetime.
• Use implicit memory. One idea is to use multiple sub-populations in solving a
dynamic optimization problem so that each sub-population can separately track
one of the optima and a new sub-population can be generated if there is a change
in the environment.
Alternatively, ideas inspired from diversity maintenance mechanisms in biology
can be used. Natural evolution has to deal with frequent environmental changes.
Thus, many genetic mechanisms have been evolved in biological systems to han-
24 1 Introduction to Optimization

dle uncertainty in the environment. One mechanism is called multiploid, which

uses multiple copies of chromosome to encode one same phenotype. Thus, dif-
ferent copies of the chromosome will ‘memorize’ the best phenotype in different
environments, and once an environmental changes occurs, the corresponding chro-
mosome will be activated and expressed, thereby quickly responding to the new
environment.
• Use explicit memory. Since it is assumed that there is certain degree of correlation
between the optima in the previous and current environment, it is useful to store
this information, in the form of solutions (the whole genome) or pieces of genomes.
Thus, these previous promising solutions or partial solutions can be added in the
new population for the new environment.
Alternatively, a model, either a probabilistic model or an associative model can be
built to explicitly store the knowledge of the promising solutions in the previous
environments to guide the search in the current environment.
• Predict the new optimum (Zhou et al., 2014). One most popular approach to
dynamic optimization is to predict the new positions of the optimum using differ-
ent models. For single-objective optimization one optimum is predicted, while for
multi-objective optimization, the Pareto front in the new environment are predicted,
based on individual solutions, the whole population or a model for describing the
Pareto front. Figure 1.12a illustrates the idea of predicting the center of a moving
Pareto front, and Fig. 1.12b is an example of predicting the moving Pareto front
using a polynomial model.
• Use multi-task optimization or knowledge transfer between the different environ-
ments (Jiang et al., 2017; Yang et al., 2020). Since the problems in different envi-
ronments are closely related, techniques such as multi-task optimization, which
was inspired from multi-task machine learning and simultaneously optimize mul-
tiple related optimization problems, are well suited for dynamic optimization. In
addition, transfer machine learning can be used for data-driven dynamic optimiza-
tion problems.
The performance of a dynamic optimization algorithm can be assessed by the best
or average performance over the generations, or the best performance averaged over
the environments. Other related measures include performance drop and convergence
speed, i.e., how fast an algorithm reacts to an environmental change. A most recent
comprehensive survey of evolutionary dynamic optimization can be found in Yazdani
et al. (2021a, 2021b).

1.4.5 Robust Optimization Over Time

Robust optimization and dynamic optimization are two very different philosophies
for handling uncertainty in optimization. While robust optimization assumes that
one single robust optimal solution can handle all uncertainty to be experienced in
the lifetime of the design, dynamic optimization hypothesizes that the algorithm is
1.4 Handling Uncertainty in Optimization 25

Fig. 1.12 a Illustration of a moving Pareto set (left), and a model predicting its center (right). b
The moving Pareto set of a dynamic multi-objective optimization problem (left), the real trajectory
of the center of the Pareto set denoted by red circles, and the predicted trajectory of the center in
different environments denoted by green diamonds and blue squares (right)

always able to timely and rapidly track the moving optimum and that the change of
solution is not subject to any additional cost. Apparently, robust optimization and
dynamic optimization are two extreme situations and none of the basic assumptions
made by them can hold in many real-world situations.
To bridge the gap between robust optimization and dynamic optimization, robust
optimization over time (ROOT) was suggested (Jin et al., 2013; Yu et al., 2010). The
main hypotheses of ROOT can be summarized as follows:
• An optimal solution should not be switched to a new solution even if an environ-
mental change is detected, so long as its performance is not worse than the worst
performance the user can tolerate.
• Switching solutions is subject to additional cost.
• Once the performance of a solution is worse than the worst performance the user
can accept, a new optimal solutions should sought. However, the algorithm will
not search for the solution having the best performance in the current environment;
instead, the algorithm searches for the optimal solution that might be acceptable
in a maximum number of environments to reduce the number of solution change.
From the above hypotheses, one can see that ROOT can be seen as a compromise
between robust optimization and dynamic optimization. In contrast to the robustness
definition in Eq. (1.37), robustness over time can be generically defined as follows:
26 1 Introduction to Optimization

0 +T +∞
t=t

FT (x) = f (x, a) p(a(t))da(t)dt, (1.46)

t=t0 −∞

where p(a(t)) is the probability density function of a(t) at time t, and T is the length
of time interval, and t0 is the given starting time. From the above definition, we can
find that ROOT not only takes into account uncertainties in the decision space and
parameter space, but also the effect of these uncertainties in the time domain.
The ROOT definition in Eq. (1.46) is very generic, which is the average perfor-
mance over the time interval T . However, it needs to be rewritten if one is interested
in finding a ROOT solution that can be used in as many environments as possible.
Given the sequence of L problems as described in (1.45), the following optimization
problem can be defined to find the robust solution over time:

maximize R = l, (1.47)
s.t. f (x, a(t)) ≤ δ, t ∈ [tc , tc + l] (1.48)

where δ is the worst performance the user can accept, tc is the starting time (the time
instant when the solution is to be adopted), and l is the number of environments the
solution will be used. In other words, the robustness is simply defined as the number
of environments the solution can be used, which is to be maximized.
Note that similar to the conventional robust optimization, there will also be a
trade-off between the robustness defined in (1.47) and the average fitness over the
whole time period [0, T ], or between the robustness and the switching cost. Thus,
ROOT can also be formulated as a bi-objective optimization problem (Huang et al.,
2017):

maximize R = l, (1.49)
minimize C = ||x − x∗ || (1.50)
s.t. f (x, a(t)) ≤ δ, t ∈ [tc , tc + l] (1.51)

where C is the cost for switching the previous optimal solution x∗ to the new optimal
solution x, which is defined to be the Euclidean distance between the two solutions.
Of course, other definitions for the switching cost are possible.
As a result, a set of Pareto optimal solutions will be found for each environment.
In practice, one of these solutions should be chosen to be implemented, e.g., the one
that maximizes the ratio between robustness and switching cost (Huang et al., 2020).
It should be pointed out that finding out the ROOT solution is non-trivial as this
requires to predict the performance of a solution in the future based on history data.
An alternative is to find out all optima in the current environment using a multi-
population strategy and then choose the ROOT solution from them according to the
properties of these solutions before and after the environmental change.
Discovering Diverse Content Through
Random Scribd Documents
These Immoralities of the Stage had by an avow'd Indulgence been
creeping into it ever since King Charles his Time; nothing that was
loose could then be too low for it: The London Cuckolds, the most
rank Play that ever succeeded,[293] was then in the highest Court-
Favour: In this almost general Corruption, Dryden, whose Plays were
more fam'd for their Wit than their Chastity, led the way, which he
fairly confesses, and endeavours to excuse in his Epilogue to the
Pilgrim, revived in 1700 for his Benefit,[294] in his declining Age and
Fortune—The following Lines of it will make good my Observation.

Perhaps the Parson[295] stretch'd a Point too far,

When with our Theatres he wag'd a War.
He tells you that this very moral Age
Receiv'd the first Infection from the Stage.
But sure, a banish'd Court, with Lewdness fraught,
The Seeds of open Vice returning brought.
Thus lodg'd (as vice by great Example thrives)
It first debauch'd the Daughters, and the Wives.
London, a fruitful Soil, yet never bore
So plentiful a Crop of Horns before.
The Poets, who must live by Courts or starve,
Were proud so good a Government to serve.
And mixing with Buffoons and Pimps profane,
Tainted the Stage for some small snip of Gain.
For they, like Harlots under Bawds profest,
Took all th' ungodly Pains, and got the least.
Thus did the thriving Malady prevail,
The Court it's Head, the Poets but the Tail.
The Sin was of our native Growth, 'tis true,
The Scandal of the Sin was wholly new.
Misses there were, but modestly conceal'd;
Whitehall the naked Venus first reveal'd.
Who standing, as at Cyprus, in her Shrine,
The Strumpet was ador'd with Rites divine, &c.
This Epilogue, and the Prologue to the same Play, written by Dryden,
I spoke myself, which not being usually done by the same Person, I
have a mind, while I think of it, to let you know on what Occasion
they both fell to my Share, and how other Actors were affected by it.
Sir John Vanbrugh, who had given some light touches of his Pen to
the Pilgrim to assist the Benefit Day of Dryden, had the Disposal of
the Parts, and I being then as an Actor in some Favour with him, he
read the Play first with me alone, and was pleased to offer me my
Choice of what I might like best for myself in it. But as the chief
Characters were not (according to my Taste) the most shining, it was
no great Self-denial in me that I desir'd he would first take care of
those who were more difficult to be pleased; I therefore only chose
for myself two short incidental Parts, that of the stuttering Cook[296]
and the mad Englishman. In which homely Characters I saw more
Matter for Delight than those that might have a better Pretence to
the Amiable: And when the Play came to be acted I was not deceiv'd
in my Choice. Sir John, upon my being contented with so little a
Share in the Entertainment, gave me the Epilogue to make up my
Mess; which being written so much above the Strain of common
Authors, I confess I was not a little pleased with. And Dryden, upon
his hearing me repeat it to him, made me a farther Compliment of
trusting me with the Prologue. This so particular Distinction was
looked upon by the Actors as something too extraordinary. But no
one was so impatiently ruffled at it as Wilks, who seldom chose soft
Words when he spoke of any thing he did not like. The most gentle
thing he said of it was, that he did not understand such Treatment;
that for his Part he look'd upon it as an Affront to all the rest of the
Company, that there shou'd be but one out of the Whole judg'd fit to
speak either a Prologue or an Epilogue! to quiet him I offer'd to
decline either in his Favour, or both, if it were equally easy to the
Author: But he was too much concern'd to accept of an Offer that
had been made to another in preference to himself, and which he
seem'd to think his best way of resenting was to contemn. But from
that time, however, he was resolv'd, to the best of his Power, never
to let the first Offer of a Prologue escape him: Which little Ambition
sometimes made him pay too dear for his Success: The Flatness of
the many miserable Prologues that by this means fell to his Lot,
seem'd wofully unequal to the few good ones he might have Reason
to triumph in.
I have given you this Fact only as a Sample of those frequent Rubs
and Impediments I met with when any Step was made to my being
distinguish'd as an Actor; and from this Incident, too, you may partly
see what occasion'd so many Prologues, after the Death of
Betterton, to fall into the Hands of one Speaker: But it is not every
Successor to a vacant Post that brings into it the Talents equal to
those of a Predecessor. To speak a good Prologue well is, in my
Opinion, one of the hardest Parts and strongest Proofs of sound
Elocution, of which, I confess, I never thought that any of the
several who attempted it shew'd themselves, by far, equal Masters to
Betterton. Betterton, in the Delivery of a good Prologue, had a
natural Gravity that gave Strength to good Sense, a temper'd Spirit
that gave Life to Wit, and a dry Reserve in his Smile that threw
Ridicule into its brightest Colours. Of these Qualities, in the speaking
of a Prologue, Booth only had the first, but attain'd not to the other
two: Wilks had Spirit, but gave too loose a Rein to it, and it was
seldom he could speak a grave and weighty Verse harmoniously: His
Accents were frequently too sharp and violent, which sometimes
occasion'd his eagerly cutting off half the Sound of Syllables that
ought to have been gently melted into the Melody of Metre: In
Verses of Humour, too, he would sometimes carry the Mimickry
farther than the hint would bear, even to a trifling Light, as if himself
were pleased to see it so glittering. In the Truth of this Criticism I
have been confirm'd by those whose Judgment I dare more
confidently rely on than my own: Wilks had many Excellencies, but if
we leave Prologue-Speaking out of the Number he will still have
enough to have made him a valuable Actor. And I only make this
Exception from them to caution others from imitating what, in his
time, they might have too implicitly admired—— But I have a Word
or two more to say concerning the Immoralities of the Stage. Our
Theatrical Writers were not only accus'd of Immorality, but
Prophaneness; many flagrant Instances of which were collected and
published by a Nonjuring Clergyman, Jeremy Collier, in his View of
the Stage, &c. about the Year 1697.[297] However just his Charge
against the Authors that then wrote for it might be, I cannot but
think his Sentence against the Stage itself is unequal; Reformation
he thinks too mild a Treatment for it, and is therefore for laying his
Ax to the Root of it: If this were to be a Rule of Judgment for
Offences of the same Nature, what might become of the Pulpit,
where many a seditious and corrupted Teacher has been known to
cover the most pernicious Doctrine with the Masque of Religion? This
puts me in mind of what the noted Jo. Hains,[298] the Comedian, a
Fellow of a wicked Wit, said upon this Occasion; who being ask'd
what could transport Mr. Collier into so blind a Zeal for a general
Suppression of the Stage, when only some particular Authors had
abus'd it? Whereas the Stage, he could not but know, was generally
allow'd, when rightly conducted, to be a delightful Method of
mending our Morals? "For that Reason, reply'd Hains: Collier is by
Profession a Moral-mender himself, and two of Trade, you know, can
never agree."[299]
 WILLIAM CONGREVE.

The Authors of the old Batchelor and of the Relapse were those
whom Collier most labour'd to convict of Immorality; to which they
severally publish'd their Reply; the first seem'd too much hurt to be
able to defend himself, and the other felt him so little that his Wit
only laugh'd at his Lashes.[300]
My first Play of the Fool in Fashion, too, being then in a Course of
Success; perhaps for that Reason only, this severe Author thought
himself oblig'd to attack it; in which I hope he has shewn more Zeal
than Justice, his greatest Charge against it is, that it sometimes uses
the Word Faith! as an Oath, in the Dialogue: But if Faith may as well
signify our given Word or Credit as our religious Belief, why might
not his Charity have taken it in the less criminal Sense?
Nevertheless, Mr. Collier's Book was upon the whole thought so
laudable a Work, that King William, soon after it was publish'd,
granted him a Nolo Prosequi when he stood answerable to the Law
for his having absolved two Criminals just before they were executed
for High Treason. And it must be farther granted that his calling our
Dramatick Writers to this strict Account had a very wholesome Effect
upon those who writ after this time. They were now a great deal
more upon their guard; Indecencies were no longer Wit; and by
Degrees the fair Sex came again to fill the Boxes on the first Day of
a new Comedy, without Fear or Censure. But the Master of the
Revels,[301] who then licens'd all Plays for the Stage, assisted this
Reformation with a more zealous Severity than ever. He would strike
out whole Scenes of a vicious or immoral Character, tho' it were
visibly shewn to be reform'd or punish'd; a severe Instance of this
kind falling upon my self may be an Excuse for my relating it: When
Richard the Third (as I alter'd it from Shakespear)[302] came from
his Hands to the Stage, he expung'd the whole first Act without
sparing a Line of it. This extraordinary Stroke of a Sic volo occasion'd
my applying to him for the small Indulgence of a Speech or two, that
the other four Acts might limp on with a little less Absurdity! no! he
had not leisure to consider what might be separately inoffensive. He
had an Objection to the whole Act, and the Reason he gave for it
was, that the Distresses of King Henry the Sixth, who is kill'd by
Richard in the first Act, would put weak People too much in mind of
King James then living in France; a notable Proof of his Zeal for the
Government![303] Those who have read either the Play or the
History, I dare say will think he strain'd hard for the Parallel. In a
Word, we were forc'd, for some few Years, to let the Play take its
Fate with only four Acts divided into five; by the Loss of so
considerable a Limb, may one not modestly suppose it was robbed
of at least a fifth Part of that Favour it afterwards met with? For tho'
this first Act was at last recovered, and made the Play whole again,
yet the Relief came too late to repay me for the Pains I had taken in
it. Nor did I ever hear that this zealous Severity of the Master of the
Revels was afterwards thought justifiable. But my good Fortune, in
Process of time, gave me an Opportunity to talk with my Oppressor
in my Turn.
The Patent granted by his Majesty King George the First to Sir
Richard Steele and his Assigns,[304] of which I was one, made us
sole Judges of what Plays might be proper for the Stage, without
submitting them to the Approbation or License of any other
particular Person. Notwithstanding which, the Master of the Revels
demanded his Fee of Forty Shillings upon our acting a new One, tho'
we had spared him the Trouble of perusing it. This occasion'd my
being deputed to him to enquire into the Right of his Demand, and
to make an amicable End of our Dispute.[305] I confess I did not
dislike the Office; and told him, according to my Instructions, That I
came not to defend even our own Right in prejudice to his; that if
our Patent had inadvertently superseded the Grant of any former
Power or Warrant whereon he might ground his Pretensions, we
would not insist upon our Broad Seal, but would readily answer his
Demands upon sight of such his Warrant, any thing in our Patent to
the contrary notwithstanding. This I had reason to think he could not
do; and when I found he made no direct Reply to my Question, I
repeated it with greater Civilities and Offers of Compliance, 'till I was
forc'd in the end to conclude with telling him, That as his Pretensions
were not back'd with any visible Instrument of Right, and as his
strongest Plea was Custom, we could not so far extend our
Complaisance as to continue his Fees upon so slender a Claim to
them: And from that Time neither our Plays or his Fees gave either
of us any farther trouble. In this Negotiation I am the bolder to think
Justice was on our Side, because the Law lately pass'd,[306] by
which the Power of Licensing Plays, &c. is given to a proper Person,
is a strong Presumption that no Law had ever given that Power to
any such Person before.
My having mentioned this Law, which so immediately affected the
Stage, inclines me to throw out a few Observations upon it: But I
must first lead you gradually thro' the Facts and natural Causes that
made such a Law necessary.
Although it had been taken for granted, from Time immemorial, that
no Company of Comedians could act Plays, &c. without the Royal
License or Protection of some legal Authority, a Theatre was,
notwithstanding, erected in Goodman's-Fields about seven Years
ago,[307] where Plays, without any such License, were acted for
some time unmolested and with Impunity. After a Year or two, this
Playhouse was thought a Nusance too near the City: Upon which the
Lord-Mayor and Aldermen petition'd the Crown to suppress it: What
Steps were taken in favour of that Petition I know not, but common
Fame seem'd to allow, from what had or had not been done in it,
that acting Plays in the said Theatre was not evidently unlawful.[308]
However, this Question of Acting without a License a little time after
came to a nearer Decision in Westminster-Hall; the Occasion of
bringing it thither was this: It happened that the Purchasers of the
Patent, to whom Mr. Booth and Myself had sold our Shares,[309]
were at variance with the Comedians that were then left to their
Government, and the Variance ended in the chief of those
Comedians deserting and setting up for themselves in the little
House in the Hay-Market, in 1733, by which Desertion the Patentees
were very much distressed and considerable Losers. Their Affairs
being in this desperate Condition, they were advis'd to put the Act of
the Twelfth of Queen Anne against Vagabonds in force against these
Deserters, then acting in the Hay-Market without License.
Accordingly, one of their chief Performers[310] was taken from the
Stage by a Justice of Peace his Warrant, and committed to Bridewell
as one within the Penalty of the said Act. When the Legality of this
Commitment was disputed in Westminster-Hall, by all I could
observe from the learned Pleadings on both Sides (for I had the
Curiosity to hear them) it did not appear to me that the Comedian so
committed was within the Description of the said Act, he being a
Housekeeper and having a Vote for the Westminster Members of
Parliament. He was discharged accordingly, and conducted through
the Hall with the Congratulations of the Crowds that attended and
wish'd well to his Cause.
The Issue of this Trial threw me at that time into a very odd
Reflexion, viz. That if acting Plays without License did not make the
Performers Vagabonds unless they wandered from their Habitations
so to do, how particular was the Case of Us three late Menaging
Actors at the Theatre-Royal, who in twenty Years before had paid
upon an Averidge at least Twenty Thousand Pounds to be protected
(as Actors) from a Law that has not since appeared to be against us.
Now, whether we might certainly have acted without any License at
all I shall not pretend to determine; but this I have of my own
Knowledge to say, That in Queen Anne's Reign the Stage was in
such Confusion, and its Affairs in such Distress, that Sir John
Vanbrugh and Mr. Congreve, after they had held it about one Year,
threw up the Menagement of it as an unprofitable Post, after which
a License for Acting was not thought worth any Gentleman's asking
for, and almost seem'd to go a begging, 'till some time after, by the
Care, Application, and Industry of three Actors, it became so
prosperous, and the Profits so considerable, that it created a new
Place, and a Sine-cure of a Thousand Pounds a Year,[311] which the
Labour of those Actors constantly paid to such Persons as had from
time to time Merit or Interest enough to get their Names inserted as
Fourth Menagers in a License with them for acting Plays, &c. a
Preferment that many a Sir Francis Wronghead would have jump'd
at.[312] But to go on with my Story. This Endeavour of the Patentees
to suppress the Comedians acting in the Hay-Market proving
ineffectual, and no Hopes of a Reunion then appearing, the Remains
of the Company left in Drury-Lane were reduced to a very low
Condition. At this time a third Purchaser, Charles Fleetwood, Esq.,
stept in; who judging the best Time to buy was when the Stock was
at the lowest Price, struck up a Bargain at once for Five Parts in Six
of the Patent;[313] and, at the same time, gave the revolted
Comedians their own Terms to return and come under his
Government in Drury-Lane, where they now continue to act at very
ample Sallaries, as I am informed, in 1738.[314] But (as I have
observ'd) the late Cause of the prosecuted Comedian having gone so
strongly in his Favour, and the House in Goodman's-Fields, too,
continuing to act with as little Authority unmolested; these so
tolerated Companies gave Encouragement to a broken Wit to collect
a fourth Company, who for some time acted Plays in the Hay-Market,
which House the united Drury-Lane Comedians had lately quitted:
This enterprising Person, I say (whom I do not chuse to name,[315]
unless it could be to his Advantage, or that it were of Importance)
had Sense enough to know that the best Plays with bad Actors
would turn but to a very poor Account; and therefore found it
necessary to give the Publick some Pieces of an extraordinary Kind,
the Poetry of which he conceiv'd ought to be so strong that the
greatest Dunce of an Actor could not spoil it: He knew, too, that as
he was in haste to get Money, it would take up less time to be
intrepidly abusive than decently entertaining; that to draw the Mob
after him he must rake the Channel[316] and pelt their Superiors;
that, to shew himself somebody, he must come up to Juvenal's
Advice and stand the Consequence:

Aude aliquid brevibus Gyaris, & carcere dignum

Si vis esse aliquis—— Juv. Sat. I.[317]
Such, then, was the mettlesome Modesty he set out with; upon this
Principle he produc'd several frank and free Farces that seem'd to
knock all Distinctions of Mankind on the Head: Religion, Laws,
Government, Priests, Judges, and Ministers, were all laid flat at the
Feet of this Herculean Satyrist! This Drawcansir in Wit,[318] that
spared neither Friend nor Foe! who to make his Poetical Fame
immortal, like another Erostratus, set Fire to his Stage by writing up
to an Act of Parliament to demolish it.[319] I shall not give the
particular Strokes of his Ingenuity a Chance to be remembred by
reciting them; it may be enough to say, in general Terms, they were
so openly flagrant, that the Wisdom of the Legislature thought it
high time to take a proper Notice of them.[320]
Having now shewn by what means there came to be four Theatres,
besides a fifth for Operas, in London, all open at the same time, and
that while they were so numerous it was evident some of them must
have starv'd unless they fed upon the Trash and Filth of Buffoonry
and Licentiousness; I now come, as I promis'd, to speak of that
necessary Law which has reduced their Number and prevents the
Repetition of such Abuses in those that remain open for the Publick
Recreation.
 CHARLOTTE CHARKE.

While this Law was in Debate a lively Spirit and uncommon

Eloquence was employ'd against it.[321] It was urg'd That one of the
greatest Goods we can enjoy is Liberty. (This we may grant to be an
incontestable Truth, without its being the least Objection to this
Law.) It was said, too, That to bring the Stage under the Restraint of
a Licenser was leading the way to an Attack upon the Liberty of the
Press. This amounts but to a Jealousy at best, which I hope and
believe all honest Englishmen have as much Reason to think a
groundless, as to fear it is a just Jealousy: For the Stage and the
Press, I shall endeavour to shew, are very different Weapons to
wound with. If a great Man could be no more injured by being
personally ridicul'd or made contemptible in a Play, than by the same
Matter only printed and read against him in a Pamphlet or the
strongest Verse; then, indeed, the Stage and the Press might
pretend to be upon an equal Foot of Liberty: But when the wide
Difference between these two Liberties comes to be explain'd and
consider'd, I dare say we shall find the Injuries from one capable of
being ten times more severe and formidable than from the other:
Let us see, at least, if the Case will not be vastly alter'd. Read what
Mr. Collier in his Defence of his Short View of the Stage, &c. Page
25, says to this Point; he sets this Difference in a clear Light. These
are his Words:
"The Satyr of a Comedian and another Poet, have a different effect
upon Reputation. A Character of Disadvantage upon the Stage,
makes a stronger Impression than elsewhere. Reading is but Hearing
at the second Hand; Now Hearing at the best, is a more languid
Conveyance than Sight. For as Horace observes,

Segnius irritant animos demissa per aurem,

Quam quæ sunt oculis subjecta fidelibus.[322]
The Eye is much more affecting, and strikes deeper into the Memory
than the Ear. Besides, Upon the Stage both the Senses are in
Conjunction. The Life of the Action fortifies the Object, and awakens
the Mind to take hold of it. Thus a dramatick Abuse is rivetted in the
Audience, a Jest is improv'd into an Argument, and Rallying grows
up into Reason: Thus a Character of Scandal becomes almost
indelible, a Man goes for a Blockhead upon Content; and he that's
made a Fool in a Play, is often made one for his Life-time. 'Tis true
he passes for such only among the prejudiced and unthinking; but
these are no inconsiderable Division of Mankind. For these Reasons,
I humbly conceive the Stage stands in need of a great deal of
Discipline and Restraint: To give them an unlimited Range, is in
effect to make them Masters of all Moral Distinctions, and to lay
Honour and Religion at their Mercy. To shew Greatness ridiculous, is
the way to lose the use, and abate the value of the Quality. Things
made little in jest, will soon be so in earnest: for Laughing and
Esteem, are seldom bestow'd on the same Object."
If this was Truth and Reason (as sure it was) forty Years ago, will it
not carry the same Conviction with it to these Days, when there
came to be a much stronger Call for a Reformation of the Stage,
than when this Author wrote against it, or perhaps than was ever
known since the English Stage had a Being? And now let us ask
another Question! Does not the general Opinion of Mankind suppose
that the Honour and Reputation of a Minister is, or ought to be, as
dear to him as his Life? Yet when the Law, in Queen Anne's Time,
had made even an unsuccessful Attempt upon the Life of a Minister
capital, could any Reason be found that the Fame and Honour of his
Character should not be under equal Protection? Was the Wound
that Guiscard gave to the late Lord Oxford, when a Minister,[323] a
greater Injury than the Theatrical Insult which was offer'd to a later
Minister, in a more valuable Part, his Character? Was it not as high
time, then, to take this dangerous Weapon of mimical Insolence and
Defamation out of the Hands of a mad Poet, as to wrest the Knife
from the lifted Hand of a Murderer? And is not that Law of a milder
Nature which prevents a Crime, than that which punishes it after it is
committed? May not one think it amazing that the Liberty of
defaming lawful Power and Dignity should have been so eloquently
contended for? or especially that this Liberty ought to triumph in a
Theatre, where the most able, the most innocent, and most upright
Person must himself be, while the Wound is given, defenceless? How
long must a Man so injur'd lie bleeding before the Pain and Anguish
of his Fame (if it suffers wrongfully) can be dispell'd? or say he had
deserv'd Reproof and publick Accusation, yet the Weight and
Greatness of his Office never can deserve it from a publick Stage,
where the lowest Malice by sawcy Parallels and abusive Inuendoes
may do every thing but name him: But alas! Liberty is so tender, so
chaste a Virgin, that it seems not to suffer her to do irreparable
Injuries with Impunity is a Violation of her! It cannot sure be a
Principle of Liberty that would turn the Stage into a Court of Enquiry,
that would let the partial Applauses of a vulgar Audience give
Sentence upon the Conduct of Authority, and put Impeachments into
the Mouth of a Harlequin? Will not every impartial Man think that
Malice, Envy, Faction, and Mis-rule, might have too much Advantage
over lawful Power, if the Range of such a Stage-Liberty were
unlimited and insisted on to be enroll'd among the glorious Rights of
an English Subject?
I remember much such another ancient Liberty, which many of the
good People of England were once extremely fond of; I mean that of
throwing Squibs and Crackers at all Spectators without Distinction
upon a Lord-Mayor's Day; but about forty Years ago a certain
Nobleman happening to have one of his Eyes burnt out by this
mischievous Merriment, it occasion'd a penal Law to prevent those
Sorts of Jests from being laugh'd at for the future: Yet I have never
heard that the most zealous Patriot ever thought such a Law was the
least Restraint upon our Liberty.
If I am ask'd why I am so voluntary a Champion for the Honour of
this Law that has limited the Number of Play-Houses, and which now
can no longer concern me as a Professor of the Stage? I reply, that it
being a Law so nearly relating to the Theatre, it seems not at all
foreign to my History to have taken notice of it; and as I have
farther promised to give the Publick a true Portrait of my Mind, I
ought fairly to let them see how far I am, or am not, a Blockhead,
when I pretend to talk of serious Matters that may be judg'd so far
above my Capacity: Nor will it in the least discompose me whether
my Observations are contemn'd or applauded. A Blockhead is not
always an unhappy Fellow, and if the World will not flatter us, we
can flatter ourselves; perhaps, too, it will be as difficult to convince
us we are in the wrong, as that you wiser Gentlemen are one Tittle
the better for your Knowledge. It is yet a Question with me whether
we weak Heads have not as much Pleasure, too, in giving our
shallow Reason a little Exercise, as those clearer Brains have that are
allow'd to dive into the deepest Doubts and Mysteries; to reflect or
form a Judgment upon remarkable things past is as delightful to me
as it is to the gravest Politician to penetrate into what is present, or
to enter into Speculations upon what is, or is not likely to come. Why
are Histories written, if all Men are not to judge of them? Therefore,
if my Reader has no more to do than I have, I have a Chance for his
being as willing to have a little more upon the same Subject as I am
to give it him.
When direct Arguments against this Bill were found too weak,
Recourse was had to dissuasive ones: It was said that this Restraint
upon the Stage would not remedy the Evil complain'd of: That a Play
refus'd to be licensed would still be printed, with double Advantage,
when it should be insinuated that it was refused for some Strokes of
Wit, &c. and would be more likely then to have its Effect among the
People. However natural this Consequence may seem, I doubt it will
be very difficult to give a printed Satyr or Libel half the Force or
Credit of an acted one. The most artful or notorious Lye or strain'd
Allusion that ever slander'd a great Man, may be read by some
People with a Smile of Contempt, or, at worst, it can impose but on
one Person at once: but when the Words of the same plausible Stuff
shall be repeated on a Theatre, the Wit of it among a Crowd of
Hearers is liable to be over-valued, and may unite and warm a whole
Body of the Malicious or Ignorant into a Plaudit; nay, the partial
Claps of only twenty ill-minded Persons among several hundreds of
silent Hearers shall, and often have been, mistaken for a general
Approbation, and frequently draw into their Party the Indifferent or
Inapprehensive, who rather than be thought not to understand the
Conceit, will laugh with the Laughers and join in the Triumph! But
alas! the quiet Reader of the same ingenious Matter can only like for
himself; and the Poison has a much slower Operation upon the Body
of a People when it is so retail'd out, than when sold to a full
Audience by wholesale. The single Reader, too, may happen to be a
sensible or unprejudiced Person; and then the merry Dose, meeting
with the Antidote of a sound Judgment, perhaps may have no
Operation at all: With such a one the Wit of the most ingenious
Satyr will only by its intrinsick Truth or Value gain upon his
Approbation; or if it be worth an Answer, a printed Falshood may
possibly be confounded by printed Proofs against it. But against
Contempt and Scandal, heighten'd and colour'd by the Skill of an
Actor ludicrously infusing it into a Multitude, there is no immediate
Defence to be made or equal Reparation to be had for it; for it would
be but a poor Satisfaction at last, after lying long patient under the
Injury, that Time only is to shew (which would probably be the Case)
that the Author of it was a desperate Indigent that did it for Bread.
How much less dangerous or offensive, then, is the written than the
acted Scandal? The Impression the Comedian gives to it is a kind of
double Stamp upon the Poet's Paper, that raises it to ten times the
intrinsick Value. Might we not strengthen this Argument, too, even
by the Eloquence that seem'd to have opposed this Law? I will say
for my self, at least, that when I came to read the printed
Arguments against it, I could scarce believe they were the same that
had amaz'd and raised such Admiration in me when they had the
Advantage of a lively Elocution, and of that Grace and Spirit which
gave Strength and Lustre to them in the Delivery!
Upon the whole; if the Stage ought ever to have been reform'd; if to
place a Power somewhere of restraining its Immoralities was not
inconsistent with the Liberties of a civiliz'd People (neither of which,
sure, any moral Man of Sense can dispute) might it not have shewn
a Spirit too poorly prejudiced, to have rejected so rational a Law only
because the Honour and Office of a Minister might happen, in some
small Measure, to be protected by it.[324]
But however little Weight there may be in the Observations I have
made upon it, I shall, for my own Part, always think them just;
unless I should live to see (which I do not expect) some future Set
of upright Ministers use their utmost Endeavours to repeal it.
And now we have seen the Consequence of what many People are
apt to contend for, Variety of Playhouses! How was it possible so
many could honestly subsist on what was fit to be seen? Their
extraordinary Number, of Course, reduc'd them to live upon the
Gratification of such Hearers as they knew would be best pleased
with publick Offence; and publick Offence, of what kind soever, will
always be a good Reason for making Laws to restrain it.
To conclude, let us now consider this Law in a quite different Light;
let us leave the political Part of it quite out of the Question; what
Advantage could either the Spectators of Plays or the Masters of
Play-houses have gain'd by its having never been made? How could
the same Stock of Plays supply four Theatres, which (without such
additional Entertainments as a Nation of common Sense ought to be
ashamed of) could not well support two? Satiety must have been the
natural Consequence of the same Plays being twice as often
repeated as now they need be; and Satiety puts an End to all Tastes
that the Mind of Man can delight in. Had therefore this Law been
made seven Years ago, I should not have parted with my Share in
the Patent under a thousand Pounds more than I received for it[325]
——So that, as far as I am able to judge, both the Publick as
Spectators, and the Patentees as Undertakers, are, or might be, in a
way of being better entertain'd and more considerable Gainers by it.
I now return to the State of the Stage, where I left it, about the Year
1697, from whence this Pursuit of its Immoralities has led me farther
than I first design'd to have follow'd it.
Ad Lalauze, sc
 CHAPTER IX.
A small Apology for writing on. The different
State of the two Companies. Wilks invited over
from Dublin. Estcourt, from the same Stage, the
Winter following. Mrs. Oldfield's first Admission
to the Theatre-Royal. Her Character. The great
Theatre in the Hay-Market built for Betterton's
Company. It Answers not their Expectation.
Some Observations upon it. A Theatrical State
Secret.
I now begin to doubt that the Gayeté du Cœur in which I first
undertook this Work may have drawn me into a more laborious
Amusement than I shall know how to away with: For though I
cannot say I have yet jaded my Vanity, it is not impossible but by
this time the most candid of my Readers may want a little Breath;
especially when they consider that all this Load I have heap'd upon
their Patience contains but seven Years of the forty three I pass'd
upon the Stage, the History of which Period I have enjoyn'd my self
to transmit to the Judgment (or Oblivion) of Posterity.[326] However,
even my Dulness will find somebody to do it right; if my Reader is an
ill-natur'd one, he will be as much pleased to find me a Dunce in my
old Age as possibly he may have been to prove me a brisk Blockhead
in my Youth: But if he has no Gall to gratify, and would (for his
simple Amusement) as well know how the Playhouses went on forty
Years ago as how they do now, I will honestly tell him the rest of my
Story as well as I can. Lest therefore the frequent Digressions that
have broke in upon it may have entangled his Memory, I must beg
leave just to throw together the Heads of what I have already given
him, that he may again recover the Clue of my Discourse.
Let him then remember, from the Year 1660 to 1682,[327] the
various Fortune of the (then) King's and Duke's two famous
Companies; their being reduced to one united; the Distinct
Characters I have given of thirteen Actors, which in the Year 1690
were the most famous then remaining of them; the Cause of their
being again divided in 1695, and the Consequences of that Division
'till 1697; from whence I shall lead them to our Second Union in——
Hold! let me see——ay, it was in that memorable Year when the two
Kingdoms of England and Scotland were made one. And I remember
a Particular that confirms me I am right in my Chronology; for the
Play of Hamlet being acted soon after, Estcourt, who then took upon
him to say any thing, added a fourth Line to Shakespear's Prologue
to the Play, in that Play which originally consisted but of three, but
Estcourt made it run thus:

For Us, and for our Tragedy,

Here stooping to your Clemency,
[This being a Year of Unity,]
We beg your Hearing patiently.[328]
This new Chronological Line coming unexpectedly upon the
Audience, was received with Applause, tho' several grave Faces
look'd a little out of Humour at it. However, by this Fact, it is plain
our Theatrical Union happen'd in 1707.[329] But to speak of it in its
Place I must go a little back again.
From 1697 to this Union both Companies went on without any
memorable Change in their Affairs, unless it were that Betterton's
People (however good in their Kind) were most of them too far
advanc'd in Years to mend; and tho' we in Drury-Lane were too
young to be excellent, we were not too old to be better. But what
will not Satiety depreciate? For though I must own and avow that in
our highest Prosperity I always thought we were greatly their
Inferiors; yet, by our good Fortune of being seen in quite new
Lights, which several new-written Plays had shewn us in, we now
began to make a considerable Stand against them. One good new
Play to a rising Company is of inconceivable Value. In Oroonoko[330]
(and why may I not name another, tho' it be my own?) in Love's last
Shift, and in the Sequel of it, the Relapse, several of our People
shew'd themselves in a new Style of Acting, in which Nature had not
as yet been seen. I cannot here forget a Misfortune that befel our
Society about this time, by the loss of a young Actor, Hildebrand
Horden,[331] who was kill'd at the Bar of the Rose-Tavern,[332] in a
frivolous, rash, accidental Quarrel; for which a late Resident at
Venice, Colonel Burgess, and several other Persons of Distinction,
took their Tryals, and were acquitted. This young Man had almost
every natural Gift that could promise an excellent Actor; he had
besides a good deal of Table-wit and Humour, with a handsome
Person, and was every Day rising into publick Favour. Before he was
bury'd, it was observable that two or three Days together several of
the Fair Sex, well dress'd, came in Masks (then frequently worn) and
some in their own Coaches, to visit this Theatrical Heroe in his
Shrowd. He was the elder Son of Dr. Horden, Minister of
Twickenham, in Middlesex. But this Misfortune was soon repair'd by
the Return of Wilks from Dublin (who upon this young Man's Death
was sent for over) and liv'd long enough among us to enjoy that
Approbation from which the other was so unhappily cut off. The
Winter following,[333] Estcourt, the famous Mimick, of whom I have
already spoken, had the same Invitation from Ireland, where he had
commenc'd Actor: His first Part here, at the Theatre-Royal, was the
Spanish Friar, in which, tho' he had remembred every Look and
Motion of the late Tony Leigh so far as to put the Spectator very
much in mind of him, yet it was visible through the whole,
notwithstanding his Exactness in the Out-lines, the true Spirit that
was to fill up the Figure was not the same, but unskilfully dawb'd on,
like a Child's Painting upon the Face of a Metzotinto: It was too plain
to the judicious that the Conception was not his own, but imprinted
in his Memory by another, of whom he only presented a dead
Likeness.[334] But these were Defects not so obvious to common
Spectators; no wonder, therefore, if by his being much sought after
in private Companies, he met with a sort of Indulgence, not to say
Partiality, for what he sometimes did upon the Stage.
In the Year 1699, Mrs. Oldfield was first taken into the House, where
she remain'd about a Twelvemonth almost a Mute[335] and
unheeded, 'till Sir John Vanbrugh, who first recommended her, gave
her the Part of Alinda in the Pilgrim revis'd. This gentle Character
happily became that want of Confidence which is inseparable from
young Beginners, who, without it, seldom arrive to any Excellence:
Notwithstanding, I own I was then so far deceiv'd in my Opinion of
her, that I thought she had little more than her Person that appear'd
necessary to the forming a good Actress; for she set out with so
extraordinary a Diffidence, that it kept her too despondingly down to
a formal, plain (not to say) flat manner of speaking. Nor could the
silver Tone of her Voice 'till after some time incline my Ear to any
Hope in her favour. But Publick Approbation is the warm Weather of
a Theatrical Plant, which will soon bring it forward to whatever
Perfection Nature has design'd it. However, Mrs. Oldfield (perhaps
for want of fresh Parts) seem'd to come but slowly forward 'till the
Year 1703.[336] Our Company that Summer acted at the Bath during
the Residence of Queen Anne at that Place. At that time it happen'd
that Mrs. Verbruggen, by reason of her last Sickness (of which she
some few Months after dy'd) was left in London; and though most of
her Parts were, of course, to be dispos'd of, yet so earnest was the
Female Scramble for them, that only one of them fell to the Share of
Mrs. Oldfield, that of Leonora in Sir Courtly Nice; a Character of good
plain Sense, but not over elegantly written. It was in this Part Mrs.
Oldfield surpris'd me into an Opinion of her having all the innate
Powers of a good Actress, though they were yet but in the Bloom of
what they promis'd. Before she had acted this Part I had so cold an
Expectation from her Abilities, that she could scarce prevail with me
to rehearse with her the Scenes she was chiefly concern'd in with Sir
Courtly, which I then acted. However, we ran them over with a
mutual Inadvertency of one another. I seem'd careless, as
concluding that any Assistance I could give her would be to little or
no purpose; and she mutter'd out her Words in a sort of mifty[337]
manner at my low Opinion of her. But when the Play came to be
acted, she had a just Occasion to triumph over the Error of my
Judgment, by the (almost) Amazement that her unexpected
Performance awak'd me to; so forward and sudden a Step into
Nature I had never seen; and what made her Performance more
valuable was, that I knew it all proceeded from her own
Understanding, untaught and unassisted by any one more
experienc'd Actor.[338] Perhaps it may not be unacceptable, if I
enlarge a little more upon the Theatrical Character of so memorable
an Actress.[339]
 SIR JOHN VANBRUGH.

Though this Part of Leonora in itself was of so little value, that when
she got more into Esteem it was one of the several she gave away to
inferior Actresses; yet it was the first (as I have observ'd) that
corrected my Judgment of her, and confirm'd me in a strong Belief
that she could not fail in very little time of being what she was
afterwards allow'd to be, the foremost Ornament of our Theatre.
Upon this unexpected Sally, then, of the Power and Disposition of so
unforeseen an Actress, it was that I again took up the two first Acts
of the Careless Husband, which I had written the Summer before,
and had thrown aside in despair of having Justice done to the
Character of Lady Betty Modish by any one Woman then among us;
Mrs. Verbruggen being now in a very declining state of Health, and
Mrs. Bracegirdle out of my Reach and engag'd in another Company:
But, as I have said, Mrs. Oldfield having thrown out such new
Proffers of a Genius, I was no longer at a loss for Support; my
Doubts were dispell'd, and I had now a new Call to finish it:
Accordingly, the Careless Husband[340] took its Fate upon the Stage
the Winter following, in 1704. Whatever favourable Reception this
Comedy has met with from the Publick, it would be unjust in me not
to place a large Share of it to the Account of Mrs. Oldfield; not only
from the uncommon Excellence of her Action, but even from her
personal manner of Conversing. There are many Sentiments in the
Character of Lady Betty Modish that I may almost say were originally
her own, or only dress'd with a little more care than when they
negligently fell from her lively Humour: Had her Birth plac'd her in a
higher Rank of Life, she had certainly appear'd in reality what in this
Play she only excellently acted, an agreeably gay Woman of Quality
a little too conscious of her natural Attractions. I have often seen her
in private Societies, where Women of the best Rank might have
borrow'd some part of her Behaviour without the least Diminution of
their Sense or Dignity. And this very Morning, where I am now
writing at the Bath, November 11, 1738, the same Words were said
of her by a Lady of Condition, whose better Judgment of her
Personal Merit in that Light has embolden'd me to repeat them. After
her Success in this Character of higher Life, all that Nature had given
her of the Actress seem'd to have risen to its full Perfection: But the
Variety of her Power could not be known 'till she was seen in variety
of Characters; which, as fast as they fell to her, she equally excell'd
in. Authors had much more from her Performance than they had
reason to hope for from what they had written for her; and none
had less than another, but as their Genius in the Parts they allotted
her was more or less elevated.
In the Wearing of her Person she was particularly fortunate; her
Figure was always improving to her Thirty-sixth Year; but her
Excellence in acting was never at a stand: And the last new
Welcome to our website – the perfect destination for book lovers and
knowledge seekers. We believe that every book holds a new world,
offering opportunities for learning, discovery, and personal growth.
That’s why we are dedicated to bringing you a diverse collection of
books, ranging from classic literature and specialized publications to
self-development guides and children's books.

More than just a book-buying platform, we strive to be a bridge

connecting you with timeless cultural and intellectual values. With an
elegant, user-friendly interface and a smart search system, you can
quickly find the books that best suit your interests. Additionally,
our special promotions and home delivery services help you save time
and fully enjoy the joy of reading.

Join us on a journey of knowledge exploration, passion nurturing, and

personal growth every day!

ebookbell.com