When I developed my first tool based on genetic algorithms, it was to replace local optimization algorithm (“Simulated Annealing”, as advertised by Numerical Recipes) as a global optimization algorithm. Now a couple of years later, I found a small book on GE that seemed on topic with what I had to implement (I relied at the time on another book that I never reviewed ; perhaps another time). Is it a good brief introduction as the book says?

Content and opinions

First of all, the book is really small, only 100 pages. At more than 50$, that’s quite a high page rate, especially when there are only 70 pages of actual content…

There are three parts in the book, the foundation, advanced optimization and learning. Let’s start with the first one.

The first part also contains the introduction. It mainly describes what evolutionary optimization means, and how to compare it to the other techniques, and a description of the other chapters. So the first interesting/technical chapter is the second one. It explains what a genetic algorithm actually is, and the way the algorithm is described is better that what I found in bigger books. There are also additional variants of algorithms used by GE (different types of recombinations for instance), other usual may be missing (like selection by tournament). As the book is about evolutionary optimization, and not GE, there is also the introduction of other algorithms, namely Particle Swarm Optimization, and Covariance Matrix Adaptation Evolution Strategies. Those were properly defined, better than most online resources as well. The third chapter is about defining the parameters of the algorithms, and this is an interesting chapter as the main topic is to allow this evolutionary algorithm to evolve the parameters during the optimization. This was really interesting to see, and I’m looking forward to implementing this in my tool.

More evolution in part 2, with using evolution to handle constraints (really funny, yet something else to try on my case), and then a variant of the Powell method mixed with evolutionary optimization. It is less useful perhaps, as it seems to be a local optimization algorithm, related to simulated annealing in some way, but with the population approach. Still interesting, but you won’t apply this as often as what you learned in the first part or the chapter before. The last chapter in this part tackles multiple objective optimization. This is really a good approach (a colleague of mine made his PhD on this kind of problems, now I understand his thesis better!) to multiple objective optimization. Usually, you have to combine them in some way, but as we use populations, we get actual borders, and possibly an infinite number of possible solutions, without having to choose which function to favor.

The last chapter is called “Learning”, but I would have called it “Applications”. The applications are model learning, but that’s about it. It is nice though that the book tackles more or less textbook functions in the first parts, and then tackles a problem that makes some actual sense (no, I don’t think the Rosenbrock cost function optimization can compare to solving an actual problem). Someone thought about what was missing in the other books!

Conclusion

In the end, the main issue is the cost of the book, although compared to other scientific books, it is not that expensive. Otherwise, I found it really clear, efficient, covering several subjects
properly. I think I found my new reference for people wanting to start evolutionary computation!