Enhancing expert system intelligentagents using game theory

An agent is generally defined as an entity capable of perceiving its environment and accomplishing a particular action without explicit instruction. Blindly taking directives without thinking is typically not intelligent, so rather a software agent is deemed intelligent if it may be characterized as situated, autonomous, flexible, and social. To behave intelligently requires decision making. Numerous fields such as economics, philosophy, and mathematics have made contributions to the realm of decision making, although none are applicable in a truly novel domain. The mathematical discipline of game theory seeks to devise the optimal strategy in strategic scenarios. Although this approach is not applicable in generalized environments it is effective in specific domains. An expert system is an intelligent agent approach to mimic human problem solving in a specific domain, and is an extension of the production system architecture. However, just as humans may take different approaches while arriving at the same solution to a problem, expert system decision making performance too is dependent upon implementation specific attempts to arrive at the desired solution quicker. These attempts are often heuristic in nature. This thesis investigates the possibility of enhancing the underlying decision making mechanism of expert systems in a more quantitative manner by incorporating the game theoretic notion of a utility measure. A high confidence factor in taking a certain action may be solidified and validated by pairing it with a high utility as well. Not only can the addition of a utility value lead to a more preferred solution, depending upon the implementation, the utility measure may also allow an expert system to arrive at a decision quicker as courses of action with greater utility as well as confidence are considered before others. As a specific example, this thesis presents an expert system which calls offensive plays in the game of basketball depending upon the observed defense. In this example, the addition of utilities enhances the ordering of the expert system rule set, and consequently outperforms the majority of random rule orderings which is equivalent to not using utility enhancement.