Towards Cooperation and Fairness in Multi-Agent Reinforcement Learning

Jasmine Jerry Aloor, Siddharth Nayak, Sydney Dolan, Victor L Qin, Hamsa Balakrishnan

Keywords: navigation, fairness, goal assignment, mappo

TL;DR: We incorporate fairness in MARL through reward tuning, where agents learn a fair assignment of goals and goal coverage in the navigation scenario using each agent’s localized observations

Abstract:

We consider the problem of fair multi-agent navigation for a large group of decentralized agents using multi-agent reinforcement learning (MARL). Previous MARL research optimizes for efficiency, safety, or both. We introduce a methodology that considers fairness over the entirety of the simulation period using a specialized reward function and decentralized goal assignment approach. Specifically, our work (1) incorporates fairness into the reward function, which leads to (2) agents learning a fair assignment of goals and (3) achieves almost perfect goal coverage in the navigation scenario using each agent’s localized observations in a decentralized manner that allows scaling to any number of agents.
We show that our work achieves higher fairness metrics when compared to other centralized goal-assignment fairness methodologies.

Submission Number: 20