Offline-Online Hybrid Reinforcement Learning Algorithm and Training-Evaluation Framework in Typical Adversarial Game Scenarios
Longfei Zhang, Zhong Liu, Xingxing Liang, Zhendu Li, Yanan Ni, Jiao Luo, Lumin Jiang · 2024
How to extract knowledge from existing historical combat data to guide combatants to make accurate and efficient decisions, reduce decision uncertainty, and improve combat effectiveness has become an important issue in winning modern warfare. In this paper, we propose a hybrid offline-online reinforcement learning (HOORL) algorithm framework based on a military chess system, which makes full use of historical empirical data and online interactive data, and combines offline reinforcement learning and online reinforcement learning methods. A series of offline-online hybrid reinforcement learning methods are derived by combining the advantages of offline and online reinforcement learning methods. To validate the performance of HOORL algorithms, this paper introduces a training and evaluation framework for hybrid reinforcement learning in typical air game scenarios. This framework, based on a military chess projection platform, supports researchers in testing and improving these algorithms.