| Game AIs as a test field for advanced AI techniques have been a popular research area. Real-Time Strategic (RTS) games particularly draw much attention for its real-time feature, vast action space, and long-term feedback which simulate complex problems in real-life. This research implemented an agent using Rainbow Deep Q Network (DQN) to microRTS as a simplified RTS simulator. The final agent achieved good performance and won over most award-winning agents with large advantage, which demonstrated Rainbow DQN a viable and effective solution in complex scenarios. Compared with the baseline DQN, Rainbow DQN had a higher learning rate and a better performance. This research further discussed the limitation of Action Component (AC) algorithms by showing that the AC agent made fewer effective actions per frame. This research added to the research conversation by first implementing Rainbow DQN to RTS games and furthering the investigation of AC algorithms, which help future research not only in RTS games but in deep reinforcement learning in complex scenarios as well. |
