Genetic-Algorithms

More than that the style is concise, engaging and consistently interesting. The explanations are practical and intuitively appealing.

I found myself jumping around in the book somewhat (I would recommend reading chapter 14 before chapter 1), but everything was explained somewhere. At the end, I felt I had a good basis not just from which to implement a grouping GA, but to consider many other variations or developments from the classic model.

The key strength of this book for me was its practical, common-sense approach that stresses understanding the problem and if it works use it, if not adapt it so it does.

Rather the spirit of GAs themselves.

What they have done is distil their learning and understanding so that others can access it more readily. I have used this book with final year project students and postgraduates.

j.e.mellor@brad.ac.uk

Evolutionary strategies are thought of as dynamical systems in the book, but these are not in general deterministic, but probabilistic in nature. The state space of the dynamical system consists of the direct product of an object parameter space, an endogenous strategy parameter set, and a collection of fitness functions. Evolution takes place in this state space via the "genetic operators", i.e. the selection, mutation, reproduction, and recombination operators. The goal of course is to find an optimum solution to the problem, and so a consideration of the convergence of the evolution strategy to this optimum must be addressed.

These issues and others, such as the differentiation between evolutionary strategies and ordinary Monte Carlo methods, are discussed in great detail in the book. The author emphasizes that the mechanism of evolutionary strategies lies in the local properties of state space, the evolutionary process being obtained by small steps in this space. He also suggests three prerequisites for the working of evolutionary algorithms, namely the evolutionary progress principle, the genetic repair hypothesis, and mutation-induced operation by recombination. The first is the statement that each change of the individuals in the state space can result in fitness gain as well as fitness loss. The second is a device employed for statistical estimation, and attempts to answer why recombinant evolution strategies are better than nonrecombinant strategies. The third is the statement that dominant recombination causes cohesion of a population and is represented by a local operator which transforms the mutations by a random sampling process.

The author makes use of differential geometry in the book to establish a theoretical framework to predict the local performance of evolution strategies. The hypersurface model is constructed as a fitness model for the calculation of progress measures, and for an elementary model of evolution dynamics. Tensor calculus is employed to study deformations of the sphere model, with the goal of obtaining useful formulae for the progress rate. A mean radius of this deformation is calculated, to serve as a substitute radius in the progress rate formulae for the sphere model.

For the case of (1+1)-selection, i.e. one parent and one offspring, where both parents and offspring are contained in the selection pool, the author derives exact integral representations for the progress rate. The quality gain for one parent and any member of offspring is also considered, and the author derives an integral expression for it using an approximation of the distribution function of the mutation-induced fitness distribution. He argues that the progress rate and the quality gain are progress measures that describe totally different aspects of the performance of evolution strategies.

The general problem of an evolution strategy with arbitrary numbers of parents and offspring is also considered. Since the distribution of parents in the parameter space is unknow, and since it changes in successive generations, this makes the analysis of the progress rate extremely difficult. The author does however derive the relations for this model in terms of a formal expression for the progress rate which is given as an integral over the distribution of a single descendant, which is generation-dependent and unknown. This distribution is approximated using Hermite polynomials and the determination of this function is then reduced to the finding of a collection of coefficients. These coefficients are functions of moments of the offspring and are estimated by the random selection process of the evolution strategy.

Recombinative evolution strategies are also studied by the author, and two special recombination types considered, namely the intermediate and dominant cases. Intermediate recombination is shown to lead to higher performance compared to nonrecombinativie strategies. The dominant case is shown to lead to mutation-induced speciation by recombination.

The author also analyzes the dynamic adaptation of the mutation strength to the local topology of the fitness landscape. Self-adaptation, which is the method for applying evolution to the adjustment of optimal strategy parameter values, is given detailed treatment for the case of one parent in terms of mean value dynamics.

In the preface it says that it is aimed a beginning graduate students, and since I have a M.Sc. in Computer Science and I just wanted to read it for fun I thought it would be for me. But I found that it uses way to many words to explain very basic things (e.g. almost a page to explain binary numbers) while many of the difficult equations just was presented without proper proof. So the book could have better if it had been cut down to a third and then supplemented with the proper proofs. So if you are a Computer Science graduate I cannot recommend this book. Given the fine books that Addison-Wesley usually publish I was quite disappointed with this one.

But if you are a student in other fields and just want an "intuitive" impression of Genetic Algorithms without the mathematical rigor it is probably good.

Chapter 1: An introduction to genetic algorithms with examples. This chapter is excellent.
Chapter 2: The mathematical theory behind genetic algorithms. This is not done very well since many of the equations isn't proven or explained properly.
Chapter 3: A Pascal program for the sample in chapter 1. This seems unneccesary since any proficient programmer easily could have implemented the program based on the information in chapter 1.
Chapter 4: The history of genetic algorithms and a number of applications all taken from research. Both seem unneccesary.
Chapter 5: An extension of the techniques presented in chapter 1. This is good.
Chapter 6-7: Introduction to machine learning. Is ok.
Chapter 8: A concluding chapter without any real new information.

The book includes many examples of problems solved with GAs, however no details are given of the implementation and the examples are presented mainly to describe the evolution of GAs.

One the plus side, book includes the code, (in Pascal) for a Simple Genetic Algorithm, (SGA), and a Simple Classifier System, (SCS). The full code is presented in the appendices, but the key sections are developed and explained in the main body.

For me, the book works two levels, the basic pieces to "play with" are presented clearly in chapters 1 and 3, and practical implementation suggestions are spread throughout the text.

By developing programs in Visual Basic, experimenting with search parameters and re-reading sections of this book - I learned something new!

It is a good book for engineers looking for practical information on genetic algorithms. The book assumes familiarity with electromagnetic theory, practical problems, and current simulation techniques.

The first three introductory chapters are solid, and universally applicable. They are a beginning overview of genetic algorithms, a more serious treatment describing pitfalls and variant techniques, and finally some rules of thumb for using the techniques.

The following chapters are detailed treatments of the application of genetic algorithms to real problems. Seven chapters are devoted to various types of antennas and antenna arrays. Additional chapters cover specific devices, such as electromagnetic filters, diffraction gratings, backscattering problems and magnetostatic devices.

This book will not teach you how to write genetic algorithm codes line by line. This book will not go into extreme depths about the theoretical limits and performance of genetic algorithms (where given, these details are not derived, simply given and explained.) There are any number of computer science and programming books which do that.

This is an engineering book: it will give many implementation details on practical problems, along with the results.

The rest of the book does dive into applying GAs to EM applications. They have a very good introduction of the motivation, problems encountered and how it was resolved. No code is available, but many have already written GAs in many programming languages.

Book assumes some knowledge in math (~calculus) and electrical engineering (my background is in physics, but it didn't handicap me). A must for anyone interested in EM design while integrating GAs or for anyone interested in applying GAs to their engineering problems.

1. Not enough step by step prodecure especially at the beginning. Mitchell is too quick to start with the math formulas. It turns out that Genetic Algorithms are fairly straight forward and easy to follow, but you have to read this book twice before you "get it" because Mitchell clouds the discussion with proofs and mathematical representations of systems. It is tough to follow.

2. Mitchell does a poor job of selecting meaningful examples to illustrate the points. A nice simple set of examples where the average person easily picture the system would have been delightful. Instead this author chooses to illustrate the Genetic Algorithms through uncommon neural networks amoung other exotic applications. I found myself struggling to understand both the example (I didn't know a thing about neural networks!) and the genetic algorithm.

When buying an Introduction type book, I expected it to be more 'down to earth'. this book is for advanced minds!

Mitchell's book is an overview of genetic algorithm analysis techniques as of 1996. The author gives a history of pre-computer evolutionary strategies and a summary of John Holland's pioneering work. A description of the basic terminology is presented and examples of problems solved using a GA (such as the prisoner's dilemma). The second chapter discusses evolving programs in Lisp and cellular automata. Also included in this chapter is a discussion of predicting dynamical systems. This was the section that has the most interest for me. Also interesting was the summary in this chapter about putting GAs into a neural network so that the ANNs could evolve.

The fifth chapter discusses when to employ a GA for maximum success. I appreciate the clearly thought out discussion of when to choose a GA for a problem. Sometimes authors of these types of books mimic the man with a hammer that thinks everything looks like a nail.

About half of the book is devoted to presenting examples of studies that have used genetic algorithms. These examples are interesting in themselves and also serve to illustrate the variety of genetic approaches that are available. The book also presents conflicting points of view of experts about which algorithms work best and why. This is helpful in combatting the impression that a beginner sometimes gets that everything is simple and all the answers are known.