Risk-attitudes, Trust, and Emergence of Coordination in Multi-agent Reinforcement Learning Systems: A Study of Independent Risk-sensitive REINFORCE

Trust facilitates collaboration and coordination in teams and is paramount to achieving optimality in the absence of direct communication and formal coordination devices. We investigate the influence of agents' risk-attitudes on trust and the emergence of coordination in multi-agent environments. To that end, we consider Independent Risk-sensitive Policy Gradient, Risk-sensitive REINFORCE, RL-agents in repeated 2-agent coordination games (Stag-Hunt). We experimentally validate our hypothesis that the agents' risk-attitudes influence coordination and collaboration by influencing the agents' learning dynamics and can lead to efficient learning of Pareto optimal policies. This suggests that risk-sensitive agents could achieve better results in multi-agent task environments.