Abstract:

Since the introduction of XCS there have been many derivative systems supporting alternative rule languages such as languages over reals, fuzzy logic, Sexpressions and even neural networks. This paper describes FOXCS, a derivative of XCS for learning rules in firstorder logic. The FOXCS system is aimed at solving tasks in modelfree, relational environments, and is generally applicable to Inductive Logic Programming (ILP) and Relational Reinforcement Learning (RRL). The system was evaluated on several benchmarking ILP tasks where it was found to perform at a level comparable to a number of wellknown ILP algorithms with regard to its predictive accuracy. This finding validates the approach of using evolutionary heuristics for discovering rules in firstorder logic under the reinforcement learning paradigm, and establishes the system as a promising alternative for RRL.
