Detecting environmental changes in cooperative multi-agent systems requires balancing speed and stability—PPR achieves this by smoothing reward signals before applying statistical drift detection, making it practical for real-world MARL deployments.
This paper addresses a critical problem in multi-agent reinforcement learning: detecting when the environment or task changes during training. The authors propose PPR (Patterns of Past Rewards), a lightweight detector that monitors agents' reward streams to identify significant shifts, enabling teams of learning agents to recognize and adapt to non-stationary environments.