Distributed decision making in checkers

The game of checkers can be played by machines running either heuristic search algorithms or complex decision making programs trained using machine learning techniques. The first approach has been used with remarkable success. The latter approach yielded encouraging results in the past, but later results were not so useful, partly because of the limitations of current machine learning algorithms. The focus of this work is the study of techniques for distributed decision making and learning by Multi-Agent DEcision Systems (MADES), by means of their application to the development of a checkers playing program. In this paper, we propose a new architecture for knowledge based systems dedicated to checkers playing. Our aim is to show how the combination of several known models for checkers playing can be integrated into a MADES, that learns how to combine individual decisions, so that the MADES plays better than any of them, without "a priori" knowledge of the quality or area of expertise of each model. In our MADES, we integrate well known search algorithms along standard machine learning algorithms. We present results that clearly show that the team as a single entity plays better than any of its components working in isolation.