Improving the accomplishment of a neural network based agent for draughts that operates in a distributed learning environment

Abstract

This article presents an extension to the system D-VisionDraughts: a draughts player agent based on a MultiLayer Perceptron Neural Network which operates in a distributed environment, and in a manner which distinguishes it from the current world champion Chinook, it learns without human supervision. The network weights are updated by Temporal Differences Methods using self-play with cloning technique. The best move is chosen by the parallel Alpha-Beta search algorithm called Young Brothers Wait Concept. The representation of the game board states is based on the NET-FEATUREMAP techniques (functions describing features inherent to Draughts game). This paper investigates the improvement obtained by D-VisionDraughts through the insertion of new features that allow a more precise representation of the board states. Further, the authors show to what extent the addition of new processors compensates the increase in training time that would be an obvious consequence of the optimization of the board state representation.