Empirical studies of default hierarchies and sequences of rules in learning classifier systems

Classifier systems are highly parallel, rule-based learning systems which are designed to continuously build and improve models of their environment based on their experience. While the basic elements of classifier systems, messages and classifiers (rules) are relatively simple, arbitrarily complex knowledge structures can be implemented by combining two basic components, default hierarchies and coupled sequences of classifiers. Classifier systems employ two learning mechanisms: (1) the bucket brigade algorithm, for allocating a credit (in the form of a single value, "strength") to existing rules based on their contributions to the system's behavior, and (2) rule discovery algorithms, including the genetic algorithm, which create rules that are plausible candidates for improving the system's knowledge base. Because strength is used both by the performance part of a classifier system, to select rules to control the system's behavior, and by the rule discovery algorithms, to control long term learning, the ability of the bucket brigade algorithm to allocate strength is critical. The results of this dissertation indicate that the standard bucket brigade algorithm must be modified in a number of ways. These changes were required primarily to reduce the disparity in strengths between rules that must cooperate if the system is to achieve and maintain acceptable levels of learning and stability. If the disparity is not controlled, the rule discovery operators lead to the spread of the higher-strength rules and the loss of useful, lower-strength rules. Results also indicate that the bid competition and message production mechanisms must be changed. Some of these changes are required to control various kinds of "parasitic" rules; others are necessary to ensure the efficient use of the limited-size message list, and yet others were necessary because of the multiple roles that strength plays (i.e., as capital to allow mistakes, to control behavior, and to bias the search for new rules). Based on these results, changes were made and the system was able to both maintain and discover structures that lead to high performance. The system was able to rapidly solve a problem that was solved earlier using an alternative, hybrid classifier system.