A multi-size convolution neural network for RTS games winner prediction

Abstract

Researches of AI planning in Real-Time Strategy (RTS) games have been widely applied to human behavior modeling and combat simulation. Winner prediction is an important research area for AI planning, which ensures the decision accuracy. In this paper, we introduce an effective architecture -- multi-size convolution neural network (MSCNN)-- into winner prediction. It can capture more feature for game states, because of the various sizes of filters in MSCNN. Experiments show that the modified evaluating algorithm can effectively improve the accuracy of winner prediction for RTS games.