• Summary
• Related Documents
  • Active Bibliography
  • Co-citation
• Version History

Abstract:

My recent research endeavours can be divided into two separate investigations. In the fall of 1998 I developed a genetic programming engine which demonstrated that multiple populations generating solutions to correlated problems exhibit coevolutionary dynamics. The results from this project were accepted for presentation at the 1999 European Conference on Artificial Life, to be held in Lausanne in September. The first section of this research summary will describe these results. Currently, I am investigating agent populations that must specialize in order to solve a given task. Specifically, I am developing an evolutionary algorithm that finds the optimal heterogeneous configuration for the population. I hope to show from this latter investigation that heterogeneous populations exhibit lower behavioural complexity than equivalent homogeneous agent populations. Heterogeneity in Problem Space In [Bongard 1999], I presented an extension to multi-population genetic algorithm research. In most multi-deme models [Cantu-Paz 1998, 1999], all demes generate candidate solutions for the same problem, or subsets of a single problem. In mymodel, each genetic programming population is responsible for generating solutions to a distinct problem. By allowing different populations to share their results, it was predicted that two populations generating solutions to similar problems would aid each other in reaching better fitness optima that they would have been able to reach in isolation. In my paper I chose symbolic regression as the problem domain; each population attempts to fit data generated by a distinct function. The data for each population was generated using a fourier series of the form y i = 5

Citations