Volume 08,Issue 03

Towards Interpretable Reinforcement Learning in Real-Time Strategy Games

Authors

Lance Bae

Abstract
In recent years, Deep Reinforcement Learning (DRL) has become an increasingly popular method of creating effective intelligence agents. DRL agents have proven to be successful, especially in the realm of games, but we as humans have difficulty understanding DRL agents’ behavior due to their complex structures. Uncovering game-playing DRL agents’ priorities and action patterns can reap valuable insights into how humans can effectively manage real-world game-like environments. One genre of games that might be of particular interest would be Real-Time Strategy (RTS) games, which involve many real-world aspects such as simultaneous management of multiple units and real-time decision making. In this paper, we introduce the method of using Decision Tree Classifiers to better understand and visualize the behaviors of DRL agents in the RTS environment, gym-µRTS.

Keyword: Explainable Reinforcement Learning, Real-Time Strategy Games.

PDF [ 752.85 Kb ] | Endnote File | XML