Efficient Reinforcement Learning for <i>StarCraft</i> by Abstract Forward Models and Transfer Learning

Injecting human knowledge is an effective way to accelerate reinforcement learning (RL). However, these methods are underexplored. This article presents our discovery that abstract forward model [thought-game (TG)] combined with transfer way. We take StarCraft II as study environment. With the help of a designed TG, agent can learn 99&#x0025; win-rate on 64&#x00D7;64 map against Level-7 built-in AI, using only 1.08 h in single commercial machine. also show TG method not restrictive it was thought be. It work roughly TGs, and be useful when environment changes. Comparing previous model-based RL, we more effective. present hypothesis gives influence different fidelity levels TG. For real games have unequal state action spaces, proposed novel XfrNet which usefulness validated while achieving 90&#x0025; cheating Level-10 AI. argue might shed light further studies efficient RL knowledge.