A machine learning approach to predict the winner in StarCraft based on influence maps

Abstract Real-Time Strategy games are very popular test beds for Artificial Intelligence (AI) researchers because they provide complex and controlled environments on which to test different AI techniques. In this paper we play the role of an external observer that tries to predict who is going to win from the events occurring during the game. In order to predict the outcome of the game, we model the game states using influence maps. Influence maps are numerical matrices representing the influence of each player’s army in the map, and they are useful for different types of spatial reasoning. We test different machine learning techniques on two different datasets of StarCraft games. The first dataset contains 4-player games in which the players are controlled by the internal game AI. The second dataset contains 2-player human games from specialized websites. We analyze the similarities and differences between both datasets and the performance of each algorithm. Finally, we perform a small experiment with expert players and conclude that our system reaches a level of precision similar to the human judges, although human judges base their predictions on a much more complex and abstract set of game features.