Multi-Agent Battlefield Game with Federated Reinforcement Learning

This paper investigates the application of federated learning (FL) in enhancing generalization and performance in unseen environments within the Battlefield multi-agent game from the PettingZoo simulator. Specifically, we address the challenge of training agents using multi-agent reinforcement learning (MARL) across diverse battlefield landscapes, each with varying wall structures and terrain. Our key contribution lies in developing a federated MARL framework that trains local models independently on distinct landscapes and aggregates them into a global model via parameter averaging. In the simulation, the performance of the global model relative to local models within a new environment has been evaluated to assess the advantages of federated learning in these settings. The experimental results demonstrated that the global model consistently outperformed local models trained in a non-federated manner, thereby validating the effectiveness of federated learning in such environments.