Motivated by the intention to increase the expressive power of learning classifier systems, we developed a new Xcs derivative, Fox-cs, where the classifier and observation languages are a subset of first order logic. We found that Fox-cs was viable at tasks in two relational task domains, poker and blocks world, which cannot be represented easily using traditional bit-string classifiers and inputs. We also found that for these tasks, the level of generality obtained by Fox-cs in the portion of population that produces optimal behaviour is consistent with Wilson's generality hypothesis.
Proceedings of the 2005 conference on Genetic and Evolutionary Computation GECCO '05. (Washington, DC 25-29 June, 2005)