Reinforcement Learning based Strategy Selection in StarCraft: Brood War

Artificial intelligence (AI) applied on real-time strategy (RTS) games, such as StarCraft has become a hot topic in recent years. Currently, many researches mainly focused on the aspects of applying reinforcement learning (RL) on the aspects, such as the building order, map analysis, macro- and micro-management, etc. This paper innovatively proposes a mechanism applying reinforcement learning on pre-designed game strategies, which can be considered as human experience and combinations of both the macro- and micro-management. In our work, game strategies are formed by many sub-strategy modules that consist a series of orderly actions corresponding to certain components and game stages. Further, by applying RL on the strategy modules, the proposed mechanism can effectively improve selection of the game strategies, and balance the AI decision and human experience. Our bot was tested against some bots selected from AIIDE 2017-2018. The results show that with the RL framework, our bot can improve the overall winning rate from 42.79% to 59.36%. Particularly, the improvements on opponent 32.06%, 25.18%, and 30.32%, against the Zerg bot, Arrakhammer, the Protoss bot, Xelnaga, the Terran bot, UAlbertaBot.