Accuracy-based learning classifier system ensembles with rule-sharing

Learning Classifier Systems (LCS) are a method of evolving compact rule-sets using reinforcement learning. This paper presents an investigation into exploiting the population-based nature of LCS for their use within highly-parallel systems, such as the National Supercomputer. In particular, the use of simple accuracy-based LCS within the ensemble machine approach is examined. Results indicate that inclusion of a rule migration mechanism inspired by Parallel Genetic Algorithms improves learning speed in comparison to an equivalent single system. A mechanism which exploits the underlying generalization mechanism of LCS is then shown further to improve performance, particularly as task complexity increases. Finally, considerably better than linear speed-up is demonstrated on a well-known benchmark task. As a result, LCS with rule-sharing was incorporated into the Supercomputer Data Mining Toolkit (EPSRC project (GR/T18455/01)) and this technique is available for use by the UK academic community, currently finding use in data mining e.g. Olympic athlete data (EP/43488/01)