Empathy in Multi-Agent Reinforcement Learning
This paper studies whether valuing other agents’ rewards can improve outcomes in multi-agent reinforcement learning. We model empathy as a reward-weighting parameter and compare fixed empathy with empathy that is learned over time using a bandit-based adaptation mechanism.
We evaluate this setup in the Prisoner’s Dilemma, the Stag Hunt, and a Renewable Resource Sharing environment. In all three settings, higher empathy clearly improves cooperation, welfare, and long-run stability. Yet when agents adapt empathy autonomously, they stabilize at intermediate levels instead of reaching the welfare-maximizing one.
Read the full paper to see how this gap emerges and why the adaptation mechanism becomes the key constraint!
