A Modified Classifier System Compaction Algorithm

Although classifier systems have displayed performance levels equaling or exceeding those of other techniques on a variety of benchmark classification problems, they usually solve those problems with a very large number of classifiers. In most cases, a large portion of the final classifier set is unneeded or wrong, with behavior masked by the correctly-functioning rules in the system. Wilson described a postprocessing procedure for reducing the number of classifiers in an XCSI classifier system while minimizing the impact of the reduction on the performance level of the system as a whole (Wilson 2001). Wilson’s procedure was designed for classifier systems that had been highly trained so that the classifiers were general in nature, and that were always correct in their classifiction of test data. In this paper, we describe some different compaction procedures that can be applied to classifier system sets that are less well-trained, that classify some instances incorrectly, or that contain classifiers that are not fully general.