Reward Design for Multi-Agent Reinforcement Learning with a Penalty Based on the Payment Mechanism

In this paper, we propose a novel method of reward design for multi-agent reinforcement learning (MARL). One the main uses MARL is building cooperative policies between self-interested agents. We take inspiration from concept mechanism game theory to modify how agents are rewarded in algorithms. defined payment that reflects negative contribution other agents’ valuation same manner as Vickrey-Clarke-Groves (VCG) mechanism. give individual agent signal consists two elements. evaluated solely on basis behavior will follow greedy and selfish policy, penalty reflect social welfare. call scheme based (RDPM). experimented with RDPM different scenarios. show can increase utility among while designs achieve far less, even basic simplistic problems. finally analyze discuss affects policy.