Opponent modeling under partial observability in starcraft with deep convolutional encoder-decoders

StarCraft, one of the most popular real-time strategy games, is a compelling environment for artificial intelligence research involving various tasks of both micro-level unit control and macro-level strategic decision making. In this study, we address an eminent problem of concern in macro-level decision making known as the “fog-of-war”, which rises from the partial observable nature of the game. Recovering information hidden under the fog can help capture advantageous high-level game dynamics, such as build orders, tactics and strategies of the opponent. Casted as a supervised learning problem, we propose a convolutional encoder-decoder architecture to predict potential counts and locations of the opponent’s units based on only partially visible and noisy state information. We visualize the model predictions on simplified grids to primarily evaluate the performance of our proposed method. Furthermore, we train an additional convolutional neural network classifier on the encoder-decoder outputs to predict the final winner of the game, as a means of demonstrating both effectiveness and applicability.