Evaluating the Effects on Monte Carlo Tree Search of Predicting Co-operative Agent Behaviour

This thesis explores the effects of including an agent-modelling strategy into Monte-Carlo Tree Search. This is to explore how the effects of such modelling might be used to increase the performance of agents in co-operative environments such as games. The research is conducted using two applications. The first is a co-operative 2-player puzzle game in which a perfect model outperforms an agent that makes the assumption the other agent plays randomly. The second application is the partially observable co-operative card game Hanabi, in which the predictor variant is able to outperform both a standard variant of MCTS and a version that assumes a fixed-strategy for the paired agents. This thesis also investigates a technique for learning player strategies off-line based on saved game logs for use in modelling.