The paper introduces ‘Risk-Sensitive Multi-Agent Reinforcement Learning in Network Aggregative Markov Games’. It discusses how classical multi-agent reinforcement learning (MARL) assumes risk neutrality and complete objectivity for agents, which falls short when representing agents with distinct subjective preferences. To address this, the authors propose a distributed sampling-based actor-critic (AC) algorithm with cumulative prospect theory (CPT) risk for network aggregative Markov games (NAMGs), called Distributed Nested CPT-AC. This method incorporates risk into the RL optimization problem, considering human economic or social preferences. Experimental results show that subjective CPT policies obtained can differ from risk-neutral ones, and agents with higher loss aversion tend to socially isolate themselves in an NAMG.
Publication date: 9 Feb 2024
Project Page: https://github.com/hafezgh/risk-sensitive-marl-namg
Paper: https://arxiv.org/pdf/2402.05906
