Self-supervised Learning for Predicting Invisible Enemy Information in StarCraft II

Insung Baek, Jinsoo Bae, Keewon Jeong, Young Jae Lee, Uk Jo, Jaehoon Kim, Seoung Bum Kim

Abstract: In real-time strategy games such as StarCraft II, players gather resources, make buildings, produce various units, and create strategies to win the game. Especially, accurately predicting enemy information is essential to victory in StarCraft II because the enemy situation is obscured by the fog of war. However, it is challenging to predict the enemy information because the situation changes over time, and various strategies are used. Also, previous studies for predicting invisible enemy information in StarCraft do not use self-supervised learning, which is extracting effective feature spaces. In this study, we propose a deep learning model combined with a contrastive self-supervised learning to predict invisible enemy information to improve the model performance. The effectiveness of the proposed method is demonstrated by quantitatively and qualitatively.