MP-Draughts: Unsupervised Learning Multi-agent System Based on MLP and Adaptive Neural Networks

This paper presents MP-Draughts: an unsupervised learning multi-agent system for Checkers whose architecture is based on adaptive and multi-layer perceptron neural networks. It is composed of agents that are trained for distinct stages of the game: one of the agents is an expert in the initial and intermediate stages of play and the remainders are experts in endgame stages. Each agent of MP-Draughts corresponds to a multi-layer perceptron neural network whose weights are updated by Temporal Difference methods. The board-states are represented by features. Each endgame agent is trained so as to be an expert in a particular profile of endgame board-state. These profiles (clusters) are mined, by means of adaptive neural networks, from a database whose elements are endgame board-states. The agent proposed here has as purpose to enhance a Kohonen SOM-based prototype of MP-Draughts previously implemented. In this prototype, the number of cluster was empirically determined and the only similarity measure investigated was the Euclidean distance. Thereby, the contributions here consist on: investigating which similarity measure proves to be more appropriate to estimate similarity among board-states that are represented by features, and improving the clustering process by automatically defining the appropriate number of clusters through adaptive neural networks. To cope with this task, in MP-Draughts a new version of adaptive neural network was proposed: the ASONDE, which modifies the SONDE neural network in such a way that it is able to deal with finite and stable databases. The accomplishment of the best architecture of MP-Draughts found in the investigations performed in this paper was evaluated in terms of the following parameters: coherence and appropriateness of the clustering process, and performance in tournaments against other unsupervised learning-based agents for Checkers.