Investigation of the Effect of “Fog of War” in the Prediction of StarCraft Strategy Using Machine Learning

Abstract

StarCraft is a well-known real-time strategy game developed by Blizzard Entertainment in 1998. One of the characteristics of this game is “fog of war,” which refers to the fact that players cannot see their opponents' regions but only their own unit. This characteristic of the game means that the information required in order to predicting the opponent's strategy is only available through “scouting.” Although the “fog of war” is one of the most important features of the game, it has not been deeply understood in the design of artificial intelligence. In this work, we propose to investigate the effect of the “fog of war” in the prediction of opponent's strategy using machine learning for human players and artificial intelligence (AI) bots. To realize this analysis, we develop a customized replay analyzer that exports the internal game events with/without the fog of war. In the experimental results, we collect replays from various sources: human vs. human, human vs. AI bots, and AI bots vs. AI bots. This systematic analysis with “fog of war” reveals the predictability of the machine-learning algorithms on different conditions and the directions for designing new artificial intelligence for the game.