Hierarchical Chunking in Classifier Systems

Proceedings of the AAAI Conference on Artificial Intelligence, 12

Volume

Search and Genetic Algorithms

Genetic Algorithms and Simulated Annealing

Two standard schemes for learning in classifier systems have been proposed in the literature: the bucket brigade algorithm (BBA) and the profit sharing plan (PSP). The BBA is a local learning scheme which requires less memory and lower peak computation than the PSP, whereas the PSP is a global learning scheme which typically achieves a clearly better performance than the BBA. This "requirement versus achievement" difference, known as the locality/globality dilemma, is addressed in this paper. A new algorithm called hierarchical chunking algorithm (HCA) is presented which aims at synthesizing the local and the global learning schemes. This algorithm offers a solution to the locality/globality dilemma for the important class of reactive classifier systems. The contents is as follows. Section 1 describes the locality/globality dilemma and motivates the necessity of its solution. Section 2 briefly introduces basic aspects of (reactive) classifier systems that are relevant to this paper. Section 3 presents the HCA. Section 4 gives an experimental comparison of the HCA, the BBA and the PSP. Section 5 concludes the paper with a discussion and an outlook on future work.