RTS Game AI Robots Winner Prediction Based on Replay Data by using Deep Learning
Huifen Li, Jie Zhou, Yu Tian · 2021
The success of the AlphaGo series of artificial intelligence (AI) programs in the game of Go has injected new vitality into AI research. The research of real-time strategy (RTS) games using deep learning algorithms provides an ideal method for AI planning, state evaluations, and human behavior research. Our work focused on winner prediction for RTS game AI robots based on states and actions information of replay dataset generated by μRTS simulator by using deep learning. Since previous winner prediction methods only encoded states information without actions information, the game information is incomplete and the accuracy of prediction is low. In the present work, we used one-hot encoding to encode both “states” elements and “actions” elements in the attributes of each sampled time point and corresponding “winner” attribute, which can achieve higher winner prediction performance. We used the encoded datasets to train (validate) and test five deep learning algorithms: CNN, MSCNN, CNP, LSTM, and BNN. The experimental results showed that all five deep learning methods can achieve an accuracy greater than 77% (often greater than 80%) using this new encoding on a task comparing with around 60% accuracy of previous algorithms. The Area Under the Receiver Operating Characteristic (ROC) Curve (AUC) value can achieve greater than 0.85 (often greater than 0.9). Finally, we sampled 20 time points from the beginning to the end of each match's replay data to illustrate the feasibility of our methods.