The Design of Innovation

Rating:
Summary: Great for GA-centered research, doubtful otherwise
Review: Genetic Algorithms, GAs, have had a brief flowering of successful application to optimization searches and their limitations have become apparent. One consequence is that a variety of alternative evolutionary computational approaches are being investigated. Another road, much less travelled, is to examine the core mechanisms of the GA concept and try to develop a second generation of improved algorithms. This is difficult work because of the very nature of the core building block theory as first proposed by John Holland. For true inovation, building blocks must be synthesized, evaluated, and combined in sucessive hierarchies, all without external intervention. David Goldberg, a stalwart Holland desciple, has been valiantly trying to extend Holland's main theorem, which applied to infinite populations and hypthetical spaces, to finite populations on real problems.

This book is actually a research monograph reporting on the results of this research. The title "The Design of Innovation" sets up a high level of expectation but the subtitle "lessons learned from and for competent GAs" is probably right. The book offers some useful insights into the internal workings of GAs and their implication for understanding true innovation. However, despite the introductory claim of an engineering approach, the book never gets around to actually showing practitioners how to apply the lessons, nor does it give direct evidence that they work as claimed (although references to recent papers which presumably demonstrate success are given).

It is perhaps ironic that the goal for GAs has been downgraded from "universal" (as first claimed by Holland) to "competent".
Goldberg's concentrates on GAs to the exclusion of other approaches that may be equally competent or even better. A further irony is the stunning admission that "for years GA practitioners have understood that commercial applications often require" combinations of GAs and other local search methods to obtain high-quality solutions in reasonable time. But if this is so, then maybe GAs aren't the best place to start in the first place.

Goldberg's ideas about the upcoming golden age of computational innovation in the last chapter are provocative. But the implication that we must await GA improvements for this to happen are a little off-putting.

In sum, this book is a well-written research monograph intended to open up further research into the heart and soul of GAs. It should be read by researchers in AI, machine learning, and related fields. However, it will not provide the immediate answers to practitioners who are now running into the limitations of GAs (and other evolutionary or general search techniques).