Player Imitation for Build Actions in a Real-Time Strategy Game

Partlan N., Madkour A, Jemmali C., Miller A.J., Holmgard C., El-Nasr M.S. (2019) Player Imitation for Build Actions in a Real-Time Strategy Game. AIIDE workshop on Artificial Intelligence for Strategy Games.


Player imitation, in which an AI agent attempts to mimic a specific player's actions, can expand the possibilities for automated playtesting and adaptive AI. Prior work on player imitation, however, has been limited to relatively simple domains. Real-time strategy (RTS) games require complex, strategic decision-making, but previous research on them has not focused on player imitation. In this study, we compare player imitation using recurrent neural networks, random forests, and a random baseline model. We test these methods using replays from the popular RTS game \textit{StarCraft}. We compare the results across methods and discuss how differences in the data sets affect performance, finding that RNNs are the most successful of the evaluated methods. We discuss takeaways for future research, including ethical considerations.