Search in Book Reviews

Species can be generated without a fitness function! If that statement did not shock and surprise you then either; 1) This book is not a book that would interest you, or 2) You spend way too much time reading genetic algorithm periodicals. Assuming that you are shocked and surprised if you are still reading, I continue. This book covers the papers that were presented at the FOGA-98 conference. These papers are mathematically, statistically and algorithmically intensive.

The collection of papers is largely disjoint as one might expect from a collection of conference papers. There has been an effort, however, to maintain some continuity, the editors tried to set up four themes and loosely group the papers under these themes:

• Genetic Algorithm (GA) dynamics,
• GAs in relation to their schemas,
• Landscape characterisation and
• GA parameter interaction

The groupings are rather effective and the editors have put the papers in an order that suggests a natural progression where appropriate. For example, the opening paper explains that species can be developed without a fitness function, a later paper (by Stephens et al) goes into a mathematical proof that proves why this happens without ever pointing that out themselves.

In general these papers are difficult to read (in the style of standard scientific papers) with frequent steps through extremely complicated mathematical equation that often 'clearly' derive one from another (well, 'clearly' if you have a doctorate in GAs!). I appreciated Prugel-Bennett's paper, which put the derivations in an appendix where they could be reviewed without impeding the flow of the paper.

As a group the papers were well presented and meticulously crafted. I take exception to Karsten and Nicole Weicker's paper which seems to make strong initial assumptions (one of which appears to be their conclusion) and then 'prove' their conclusion. The paper seems to assume that the fitness of a population can be no better than the parent population (if this is true, then GAs aren't useful) and using this assumption demonstrate how various algorithms are stymied by local optima. Unless I have misread the paper, this was a foregone conclusion (any parent in the proximity of a local optimum will have children who can do no better than the local optimum).

This book is an excellent (though very detailed) accompaniment to more easily understood books on similar topics such as Kauffman'sThe Origins of Order. I highly recommend this book for what it is and bearing in mind that this is not a book for a general audience.