Investigating Reinforcement Learning and Evolutionary Computation for Games with Stochasticity and Incomplete Information

Abstract

This thesis presents an application of reinforcement learning and evolutionary computation for solving complex games with incomplete information and stochasticity. Although there has been significant recent progress on AI game players, traditional deep reinforcement learning methods have mainly shown success in games with simpler properties. In this thesis, we evaluate two deep reinforcement learning methods: policy gradient and evolutionary strategies for training the neural network behind the AI players for Ticket to Ride, a complex strategic board game. By comparing AI players’ performance and policies with existing heuristics players, we show that the AI players learn well under both training algorithms. Furthermore, the results indicate that training the AI players under the complete information game environment has a positive influence on their performance under the incomplete information game environment as well.