Genetic programming for subjective fitness function identification

Abstract

This work addresses the common problem of modeling fitness functions for Interactive Evolutionary Systems. Such systems are necessarily slow because they need human interaction for the fundamental task of fitness allocation. The research presented here demonstrates that Genetic Programming can be used to learn subjective fitness functions from human subjects, using historical data from an Interactive Evolutionary system for producing pleasing drum patterns. The results indicate that GP is capable of performing symbolic regression even when the number of training cases is substantially less than the number of inputs. © Springer-Verlag 2004.