Learning Map-Independent Evaluation Functions for Real-Time Strategy Games

Real-time strategy (RTS) games have drawn great attention in the AI research community, for they offer a challenging and rich testbed for both machine learning and AI techniques. Due to their enormous state spaces and possible map configurations, learning good and generalizable representations for machine learning is crucial to build agents that can perform well in complex RTS games. In this paper we present a convolutional neural network approach to learn an evaluation function that focuses on learning general features that are independent of the map configuration or size. We first train and evaluate the network on a winner prediction task on a dataset collected with a small set of maps with a fixed size. Then we evaluate the network's generalizability to three set of larger maps. by using it as an evaluation function in the context of Monte Carlo Tree Search. Our results show that the presented architecture can successfully capture general and map-independent features applicable to more complex RTS situations.