XCS and the Monk's Problems

It has been known for some time that Learning Classifier Systems (Holland, 1986) have potential for application as Data Mining tools. Parodi and Bonelli (1990) applied the Boole LCS (Wilson, 1985) to a Lymphography data set and reported 82% classification rates. More recent work, such as GA-Miner (Flockhart, 1995) has sought to extend the application of LCS to larger commercial data sets, introducing more complex attribute encoding techniques, static niching, and hybrid genetic operators in order to address the problems presented by large search spaces. Despite these results, the traditional LCS formulation has shown itself to be unreliable in the formation of accurate optimal generalisations, which are vital for the reduction of results to a human readable form. XCS (Wilson, 1995, 1998) has been shown to be capable of generating a complete and optimally accurate mapping of a test environment (Kovacs, 1996) and therefore presents a new opportunity for the application of Learning Classifier Systems to Data Mining. As part of a continuing research effort this paper presents some first results in the application of XCS to a Data Mining task. It demonstrates that XCS is able to produce a classification performance and rule set which exceeds the performance of most current Machine Learning techniques when applied to the Monk's problems (Thrun, 1991).