Reinforcement learning applied to the game of Abalone

Abstract

This master’s thesis applies artificial intelligence methods to ABALONE, a two-playerstrictly competitive game of perfect information. It assumes no prior knowledge ofartificial intelligence. Reinforcement learning (RL) is the problem faced by an agentthat learns behaviour through trial-and error interactions with a dynamic environment.This work gives a firm mathematical foundation of RL in combination with neuralnetworks. A feedforward neural network is conceived that is able to learn on its own toplay ABALONE, without the aid of an intelligent teacher. Experiments are conducted todetermine its level of play and the influence of various learning parameters. ‘The workis completed by the development and analysis of advanced game-searching algorithms.