Opponent Strategy Recognition In Real Time Strategy Game Using Deep Feature Fusion Neural Network

This paper focuses on the task of opponent strategies recognition in Real-time Strategy (RTS) Game, which aims to predict opponent strategies by modeling the observable environmental information. It is a very challenging task due to two folds. (1) In RTS game, the information is imperfect due to the fog of war; and (2) the action and environment spaces of RTS game are too vast and complex to be modeled. This task is also significative since opponent strategies recognition is a crucial component of creating high-level AI system that can defeat high-level human players in RTS game. Most previous approaches focus on predicting tech tree, building order and strategies through game logs, where perfect information is utilized. Accordingly, these prediction methods cannot be applied to real AI systems confronting the fog of war. Furthermore, conventional approaches use machine learning techniques such as Hidden Markov Model (HMM) and Bayesian network, which is difficult to deal with higher-dimensional state spaces. Besides, the hand-crafted features are commonly used instead of high-dimensional feature of the complex environment, which leads to loss of information of the environment. To address these problems, we propose a deep feature fusion neural network to handle the above imperfect and complex information of the environment for opponent strategies recognition in RTS game. We test our method on the canonical RTS game, i.e., Starcraft II, and promising performance has been obtained.