Supervised Learning of StarCraft 2 Game Prediction Using RNN

Open Access
Lu, Sen
Area of Honors:
Computer Science
Bachelor of Science
Document Type:
Thesis Supervisors:
  • Daniel Kifer, Thesis Supervisor
  • Rebecca Jane Passonneau, Honors Advisor
  • Supervised Learning
  • Deep Learning
  • RTS Game
  • StarCraft 2
  • RNN
  • GRU
  • Keras
  • data generator
  • partial observation
Recent achievements by AlphaGo have inspired many interests in training sophisticated agents to excel in more challenging games. This paper aims to aid the reinforcement learning agent by providing it a more accurate predictor of game outcome which can effectively improve the reward function of the agent. The supervised predictor trained uses deep learning techniques, especially RNNs, on the replay data. While this is for StarCraft 2 replays specifically, the analysis and learning model of replay observations can extend to other video games with little modifications.