Product Description

This book is the only source that provides comprehensive, current, and correct information on problem solving using modern heuristics. It covers classic methods of optimization, including dynamic programming, the simplex method, and gradient techniques, as well as recent innovations such as simulated annealing, tabu search, and evolutionary computation. Integrated into the discourse is a series of problems and puzzles to challenge the reader. The book is written in a lively, engaging style and is intended for students and practitioners alike. Anyone who reads and understands the material in the book will be armed with the most powerful problem solving tools currently known. This second edition contains two new chapters, one on coevolutionary systems and one on multicriterial decision-making. Also some new puzzles are added and various subchapters are revised.

Product Details

• Amazon Sales Rank: #208632 in Books
• Published on: 2004-09-21
• Original language: English
• Number of items: 1
• Binding: Hardcover
• 554 pages

Customer Reviews

Wide, deep and good fun. Well worth reading.
Although a text book, this book is fun. It is both a good review of evolutionary algorithms and an excellent introduction to problem solving.

Algorithms are not in general given in full and there is no source code, so you will still have to put some effort in if you have a real problem to solve. However this is one of the points of the book: real problems are complex, you can't just use a recipe.

Coverage of other search methods such as neural nets is not so extensive, but there is reasonable coverage of more traditional methods.

On top of all this, there are numerous problem solving problems. These are all good fun and need nothing more than paper and pen.

A good book for the train and for the library. And there arn't many of those.

Underwhelming
Having read a couple of very positive reviews, I was looking forward to this, but now that it is here, and I've had a chance to read it, I'm very disappointed. It strikes me as pretty shallow, and it definitely takes the name of Georg Polya in vain. (Michaelewicz also writes 'business-oriented' books on decision support - you can tell). It certainly has little, conceptually, to do with Polya's 'How to solve it' (in fact, given the complete lack of any formal theoretical development, the authors are lucky that the man is safely dead). A more accurate title would be 'a bunch of stuff on optimisation, mostly about genetic algorithms and traveling salesman problems, but with a bit on neural nets and fuzzy logic thrown in'. These three technologies used to get sexy articles in the popular computer press about 10 to 20 years ago. It is interesting why the three always seem to crop up together, but they do - or at least they did.

Anyway, the core agenda, which is not heuristics, does poke out at various points. On page 190 there is a revealing passage bout the elusive 'Holy Grail' of 'a perfect evolutionary algorithm for the TSP [Travelling Salesman Problem]'. Now, the world in general would be fascinated by a polynomial solution to the TSP, but the world in general - sorry to say - doesn't actually give a toss if that solution is evolutionary.*

As I said, I was unhappy about the complete lack of real theoretical background which would put any of the discussed methods in perspective/context. The discussion of simulated annealing, for instance, is absent any of the underlying (and powerful) intuitions from statistical physics which, if nothing else, makes the technique much richer, and not conceptually comparable to, tabu search, with which it is discussed in parallel. As far as I can see, the latter is an isolated hack - empirically it may be effective in some applications, but it is not part of a larger conceptual framework. At least if it is not an isolated hack, then the authors provide no evidence - I note that it gets all of three lines in Russell and Norvig.

More seriously, there isn't even any well-founded discussion of mathematical models of evolution. All that I could find was essentially a citation - not even a discussion - of the 'no free lunch' theorems. No Maynard-Smith, no Kondrashov, not even Hopfield's '78 paper - though other later stuff by Hopfield is cited. And without this - and without a lot of other formal theory stuff as well, if I were being honest - there is no hope of a methodological framework. After all, computers are, unavoidably, formal machines. In the end, the evolutionary models that are discussed are not a lot more than a bunch of gadgets, and the authors are reduced to saying that for your own problems, you are going to have to think up your own gadgets. This, to be blunt, is why evolutionary computing remains a niche research area.

A separate problem is the lack of any perspective w.r.t. other, today more commonly used, technologies that address similar problems. The discussion of neural networks and pattern classification does not mention, e.g., that the benchmark classifier technology today is vector support machines, which substantially outperform neural networks. Or that the standard technology for exploring complex function spaces is Monte-Carlo analysis (strictly, a lot of the techniques that are discussed in the book _are_ actually Monte-carlo methods of one sort or another). This does not mean that the methods discussed are _not_ interesting, just that without a perspective, it is difficult to say whether they are appropriate tools for a job.

In the end it was not clear to me who this book is really aimed at. It is certainly not aimed at me. Senior undergraduates in something like operations research might be a target, but I personally would not use it for a CS or applied math class. And I cannot, honestly, forsee it being a lot of use to me outside the classroom.

*Which anyway, a priori, seems unlikely, since idealised recombinatory evolutionary strategies show rapid information gain (N^1/2) in the size of the genome in suboptimal situations, but near the optimum, parthenogenesis with mutation is a better strategy.