Opponent Resource Prediction in StarCraft Using Imperfect Information

Abstract

The real-time strategy (RTS) game StarCraft has recently become a focus of research on game AI. A major challenge in RTS gameplay is making decisions using imperfect information about the opponent's state and actions. One approach that has proven rewarding is to apply machine learning techniques to replays of games between skilled human players. We consider the problem of estimating the number of resources gathered by the opponent during a StarCraft match. We introduce and evaluate two techniques for opponent resource prediction using supervised learning on match replays. Our first method uses multiple linear regression on observable features of the game state. Our second method uses naïve Bayes classification to form imprecise but accurate predictions.