Inverse Emergence: Using Multi-Agent RL to Model and Control Unintended Collective Behaviors

While much MARL research seeks better cooperation or control, large-agent systems (e.g., city-scale resource allocation, as in Wang et al., 2024) are prone to surprising, often undesirable, emergent behaviors—think traffic jams, market crashes, or coordination breakdowns. This idea proposes to use inverse RL (IRL) to “reverse-engineer” the collective policies that give rise to such anomalies. By observing the system when anomalies occur, the IRL module infers the implicit objectives or reward structures that led to the behavior. The system can then (1) predict when such behaviors are likely to occur, (2) design counter-policies or interventions to steer the system away from undesirable states, and (3) provide insights for policy makers (e.g., city planners) about the true motivations driving collective agent behavior. This is a step beyond masking or fixing anomalies; it’s about understanding and controlling them at a systemic level—potentially opening new frontiers in robust, explainable multi-agent systems.

References:

1. DyPS: Dynamic Parameter Sharing in Multi-Agent Reinforcement Learning for Spatio-Temporal Resource Allocation. Jingwei Wang, Qianyue Hao, Wenzhen Huang, Xiaochen Fan, Zhentao Tang, Bin Wang, Jianye Hao, Yong Li (2024). Knowledge Discovery and Data Mining.