Assessing the Cognitive Abilities of Alternate Learning Classifier System Architectures

摘要

Since its inception in the 1960s, the Genetic Algorithm (GA) framework for solving complex problems has been simultaneously intensely studied and deployed. Despite wide-ranging practical successes in engineering, manufacturing, applied, and social science domains, developing GA-based systems has been more art than science. Consequently some researchers have attempted to build and test theories and models for robust GA design. Given this attention to 'pure' Genetic Algorithm research and implementations, progress on a subsequent GA-based framework called Learning Classifier Systems (LCS) lay dormant until the late 1990s. Stalwarts in GA/LCS research have opined that to further advance the field and facilitate theory formation, a broad study of LCSs, particularly one that focuses on their cognitive aspects, is needed. I wish to contribute to this theory building effort by examining, using simulation modeling and analyses, how alternative LCS architectures learn to cope with other artificial entities in challenging, artificial environments created using variants of the Iterated Prisoners Dilemma (PD) Tournament setting.. The use of competing entities in this setting may be likened to a number of practical applications in which different agents must negotiate or compete with each other. One possible application is the use of computer-based agents in negotiations in a buying- selling situation. In such an environment, a buyer's agent must attempt to discern the seller's negotiation pattern, and then use this information to accomplish its objective. In this example, an LCS- based agent could be used in repeated encounters with the seller to improve its performance with regard to a measure of interest such as price, quantity or delivery time.