Emergent Collective Behavior in Multi-Agent Reinforcement Learning with Diverse Attitudes

We study the emergence of collective behavior in systems of interacting reinforcement learning (RL) agents with heterogeneous reward structures. Using RL simulations inspired by the videogame Overcooked, we explore cooperative and competitive dynamics in a shared environment. In parallel, we extend the teacher–student framework to multi-agent settings with analogous reward structures to study such collective dynamics in an analytically tractable setting. In both approaches, we define agents with distinct attitudes—individualistic, competitive, or cooperative—implemented through different reward functions. Our simulations and analytical results show that combinations of these attitudes lead to qualitatively distinct forms of collective behavior. Moreover, we find that the same individual attitude can give rise to different policies depending on the attitudes of others. These preliminary results suggest a promising route toward identifying fundamental principles that govern the emergence of collective intelligence and coordination from individual learning rules.