Curriculum Interleaved Online Behavior Cloning for Complex Reinforcement Learning Applications
Michael Möbius, Kai Fischer, Daniel Kallfass, Stefan Göricke, Thomas Doll · 2024
This paper introduces Curriculum Interleaved Online Behavior Cloning (IOBC) as an approach to train agents for military operations, addressing not only the challenges posed by complex and dynamic combat scenarios but also how military doctrines and strategies are transferred to these agents. It highlights the limitations of traditional reinforcement learning (RL) methods and proposes interleaved online behavior cloning in combination with curriculum learning as a solution to enhance RL agent training. By leveraging rule-based agents for guidance during training, IOBC accelerates learning and improves the RL agent's performance, particularly in early stages of training and complex scenarios. The study conducted experiments using ReLeGSim, a reinforcement learning-focused simulation environment, demonstrating the effectiveness of our method in enhancing agent performance and scalability. Results indicate that IOBC significantly outperforms RL agents without guidance, providing a stable foundation for learning in challenging environments. These findings underscore the potential of IOBC in real-world military applications.