Making and Acting on Predictions in StarCraft: Brood War

Abstract

Making predictions and strategic decisions is not just a problem in real life, but also in the complex environments of most real-time strategy games. Due to the lack of complete information, such games can be used as testbeds for AI research in this area. This thesis presents how an AI agent, in the game StarCraft, can make use of a Bayesian network to predict the state of the opposing player's technology tree. It is also shown how the predictions can be used to assist the AI agent in making strategic decisions. Bayesian networks are used for the predictions, and the agent's army composition is generated dynamically based on the predictions and the observed opponent units. The agent is tested against StarCraft's built-in AI. The results show that it is possible to accurately predict the state of the opponent's technology tree, and that the predictions have a positive e ect on the AI agent's win rate.