Evolutionary Search for Executable Solutions

Richard K. Belew

Computer Science & Engr. Dept.
Univ. California -- San Diego

Presented to CSE Faculty Proseminar, 26 Nov 97


The Genetic Algorithm (GA) is now widely used in a range of contexts, from drug design to the inspection of silicon wafers. In most current applications, the problem is posed as one of "function optimization," where a very large, high- dimensional space of is searched for the parameter values which globally optimizes some objective function.

This talk will sketch some of the techniques as we attempt to use the GA for an even more difficult search: finding PROGRAMS. We know from standard results (like the Halting and Busy Beaver problems) that many tests we might have for programs are uncomputable. On the other hand, a number of recent experiments have shown that some functional solutions can be discovered by "genetic programming" techniques. As time allows, I will survey three techniques explored by our lab towards this goal: evolving neural networks; evolving grammars; and coevolution.

Overheads used in the talk (Postscript, 242K, compressed)