PPR Detects Task Changes in Multi-Agent RL From Reward Streams Alone, Trading Detection Speed for Fewer False Alarms

arXiv 2609.05298·low signalXBlueskyLinkedInCopy link

arXiv 2609.05298 studies online change-point detection for cooperative MARL using only reward-derived signals, since agents must notice the environment or objective changed before they can adapt. Patterns of Past Rewards is an algorithm-agnostic detector that smooths return streams, highlights recent changes and applies a statistical drift detector. In a custom Speaker-Listener environment built on the Multi-Agent Particle Environment under two controlled non-stationarity scenarios, a smoothed-return baseline detected earlier but fired many repeated alarms while raw returns often missed the shift entirely; PPR sits between them.