Developing a Reinforcement Learning Agent for the Game of Checkers

Abstract

The aim of the following paper was to develop, test and evaluate a simple self-learning agent for the game of Checkers based on reinforcement learning and neural networks. The approach followed in this work is rather simple and based on a single deep neural network which is used to evaluate the board states and to choose the next best move for the agent. During the training phases the neural network is trained using a reward system based on different criteria derived from the prior and the current board state that resulted from the last action taken. The neural network takes a state-action pair, consisting of the current state and a possible move option and predicts the total reward that can be expected. This way all possible move options are evaluated and the one with the highest value is chosen as the next move.