Nash Equilibria in Multi-Agent Swarms


In various settings in nature or robotics, swarms offer various benefits as a structure that can be joined easily and locally but still offers more resilience or efficiency at performing certain tasks. When these benefits are rewarded accordingly, even purely self-interested multi-agent reinforcement learning systems will thus learn to form swarms for each individual’s benefit. In this work we show, however, that under certain conditions swarms also pose Nash equilibria when interpreting the agents’ given task as multi-player game. We show that these conditions can be achieved by altering the area size (while allowing individual action choices) in a setting known from literature. We conclude that aside from offering valuable benefits to rational agents, swarms may also form due to pressuring deviants from swarming behavior into joining the swarm as is typical for Nash equilibria in social dilemmas.

12th International Conference on Agents and Artificial Intelligence (ICAART 2020)