Reinforcement Learning in Board Games

This project investigates the application of the T D(*) reinforcement learning algorithm and neural networks to the problem of producing an agent that can play board games. It provides a survey of the progress that has been made in this area over the last decade and extends this by suggesting some new possibilities for improvements (based upon theoretical and past empirical evidence). This includes the identification and a formalization (for the first time) of key game properties that are important for TD-Learning and a discussion of different methods of generate training data. Also included is the development of a TD-learning game system (including a game-independent benchmarking engine) which is capable of learning any zero-sum two-player board game. The primary purpose of the development of this system is to allow potential improvements of the system to be tested and compared in a standardized fashion. Experiments have been conduct with this system using the games Tic-Tac-Toe and Connect 4 to examine a number of different potential improvements.