Genetic algorithms refer to computer programs that 'evolve' in ways similar to biological organisms. 'Natural selection' specifies the features of the solution to look for, strings of binary numbers (or other similar structures) are mated, with the combination of strings containing partial solutions often producing the most 'fit' results. Generation after generation of this process continues towards the 'evolution' of the desired features. Although this reference is quite long, it is quite readable, and can be shortened significantly by omitting a number of subsections as well as chapters not essential to the core concepts, as well as the detailed appendices. This reference shows that a variety of problems from different fields can be solved in terms of a computer program, of which genetic programming can be the means to find one or more such valid computer programs. It is relevant in that genetic programming is another way to effect computation, as well as providing insight with respect to evolution in nature.

I bought Genetic Programming (GP) I & II many years ago. While I have yet to find a useful application of Koza'a work to my problems, I think many of the ideas he introduces make significant ground in many AI areas. The GP community grows every year and this is the book that started it all. Definately worth a read.

Puntos a favor: - descripción original de la GP - muchos ejemplos de su aplicabilidad - fácil de comprender

Puntos en contra: - código fuente en un ápendice - código fuente de solo 3 ejemplos simples - código en LISP, no usa CLOS - sin indicaciones de como portarlo a C++ - no incluye código de funciones autom. definidas - se concentra excesivamente en mediciones estadísticas y menos en la técnica de resolver problemas

The short history of computer science as a discipline has had two major concerns: the production of programs that are provably efficient, and the production of programs that are provably correct. "Genetic Programming" is, possibly, the beginning of a third stream in CS, the production of programs that are possibly neither efficient nor correct, but "fit" to perform a given task. A strange idea to computer scientists, perhaps, but consider the analogy with living creatures. Is a shark, a bee, or a turtle either "efficient" or "correct"? Perhaps, perhaps not; there doesn't seem to be a way to measure these concepts for something as complex as a living species. But they are "fit." They've been successful, as species, in their respective ecological niches for millions of years. Koza's big idea is the automatic generation of programs via mutation and selection, by analogy with living systems, and he's written a big book to go with the big idea (819 pages). Demonstrating creation of non-trivial programs by means of simulated mutation & selection is a major accomplishment. I'd rate the promise of this line of research as high, given that compute power becomes cheaper every year while human brain power becomes more expensive. Also, natural systems are resilient and adaptive to changes in the environment, while man-made software systems are all too fragile. This observation leads to the hope that "fit" programs may increase the robustness of the the computer networks on which so much now depends. One quibble: there is a thin book inside this fat book, trying to get out. The thin book would make the research more accessible to the average practicing programmer. Until such a "reader's edition" comes out, "Genetic Programming" is a unique resource volume.