GP Genome Representation

Relevant papers from the Symposium Working Notes:

Evolving a Team. Thomas Haynes, Sandip Sen, Dale Schoenefeld and Roger Wainwright

GPRobots and GPTeams - Competition, Co-evolution and Co-operation in Genetic Programming. Conor Ryan

Coevolution may hit on an upper limit of utility. We expect an arms race, with both populations improving beyond what independent evolution against fixed test cases could provide. I have seen situations where one population learned a strategy that the other could not cope with. This was due in part to the lack of power in the language, i.e. function and terminal sets. I can see coevolution carrying on in an arms race until such a situation occurs. I can also see coevolution eventually degrading into a cyclic struggle for power, with two strategies in a population taking turns being dormant. Only becoming active as the other side flips through its two strategies. {Thomas Haynes}

Jordan Pollacks idea about competition between individuals - scoring individuals relative to the test cases they solved that the other DIDN'T solve. This was the best - and one of the simplest - ideas I've heard for a long time. {Conor Ryan} [Competitively coadapting the distribution of test cases has a similar effect as far as "harder" test cases becoming more significant in the fitness measure. {Eric Siegel}]

The population cooperation of classifier systems may have a similar effect to competitive GP/GA systems, since individuals are rewarded more strongly for what others do wrong. {Lee Spector}