Improving NetFeatureMap-Based Representation through Frequent Pattern Mining in a Specialized Database

The adequate representation of states in the construction of intelligent agents is fundamental for allowing them to achieve a satisfactory performance, principally for those that actuate in a competitive environment that possesses a high state space. One particular type of representation that is very appropriate for these situations is the NetFeatureMap, which describes by means of features the relevant aspects that are inherent to the environment where the agent actuates. In renowned intelligent agents, such features are manually selected, which certainly leads to inadequate choices. Thus, investigating adequate approaches that perform automatic selection of these features becomes a crucial task. In this way, the main contribution of this paper is to propose a new approach that automatically selects appropriate features based on the frequency at which they occur in the states explored by the agent in the course of its acting over the environment. Such an approach is based on Frequent Pattern Mining. It is interesting to point out that there also exist Genetic Algorithms-based approaches that successfully cope with the same task. Unlike Genetic Algorithms that use heuristic functions to select the features, the present proposal uses real data contained in a specialized database for performing this task. Under the intent of investigating the efficacy of such a proposal, the authors utilize the domain of Checkers player agents as their case study, since they operate in a competitive environment with a very wide state space. This investigation is performed by means of tournaments in which agents whose features are selected by the approach proposed herein face others whose features are selected either manually or by Genetic Algorithms. The superior performance of the Frequent-Pattern-based agents in the tournaments proves the efficacy of the present proposal.