KSEF R&D Excellence: Assessing the Evolutionary Behavior of Alternate Learning Classifier System Architectures in the Iterated Prisoner's Dilemma Environment

PROJECT SUMMARY ASSESSING THE EVOLUTIONARY BEHAVIOR OF ALTERNATE LEARNING CLASSIFIER SYSTEM ARCHITECTURES IN THE ITERATED PRISONER'S DILEMMA ENVIRONMENT Principal Investigator: Ramakrishnan (Ram) Pakath, Ph.D. Affiliation: DSIS Area, School of Management Suite 425, C. M. Gatton College of Business and Economics University of Kentucky Lexington, KY 40506-0034. Email: [email protected];[email protected] Phone: 859-257-4319 (personal); 859-257-3080 (area secretary) Fax: 859-257-8031. We seek to advance extant knowledge on machine learning systems by further examining a currently popular mechanism for adaptive learning called XCS, a descendant of the Learning Classifier System (LCS) family. Learning Classifier Systems, based on a machine learning paradigm called Genetic Algorithms, evolve syntactically simple string rules called classifiers to guide system performance in unknown, arbitrary environments. XCS builds on this model, differing primarily in its assessment of classifier utility by focusing on payoff prediction accuracy in lieu of payoff magnitude. Under appropriate circumstances, such as solving Boolean multiplexer functions and guiding animats through maze-like environments, specific XCS implementations evolve an "optima'" set of classifiers that better "displays" the evolved knowledge. While such investigations are indeed significant in and of themselves, they involve building artificial systems capable of rational thought in fairly narrow settings. We seek to further examine XCS's performance in artificial situations with broader commercial/social parallels where the setting is non- Markov and asymmetric and one where the system can sometimes benefit from irrational behavior. The Iterated Prisoner's Dilemma (IPD) game-playing scenario offers us a convenient means to study such behavior, as well as other features. Whereas only limited research has been performed on the "traditional" LCS model's evolutionary behavior in IPD settings, even less is known regarding XCS in this romplex environment, particularly relating to its hypothesized advantages over LCS. In this research, we progressively and systematically perturb a .conventional" IPD-playing LCS system until it results in a fullfledged XCS system, comparing and contrasting each LCS perturbation with XCS in terms of a number of performance measures. Our intent is to examine the XCS paradigm to understand how it better copes with a given IPD situation (if it, indeed, does) than the LCS perturbations examined. Key Words/Phrases: Genetic Algorithms, Learning Classifier Systems, XCS, Machine Learning, Iterated Prisoner's Dilemma