Emergent Prosociality in Multi-Agent Games Through Gifting.

摘要: Coordination is often critical to forming prosocial behaviors -- that increase the overall sum of rewards received by all agents in a multi-agent game. However, state art reinforcement learning algorithms suffer from converging socially less desirable equilibria when multiple exist. Previous works address this challenge with explicit reward shaping, which requires strong assumption can be forced prosocial. We propose using restrictive peer-rewarding mechanism, gifting, guides toward more while allowing remain selfish and decentralized. Gifting allows each agent give some their other agents. employ theoretical framework captures benefit gifting equilibrium characterizing equilibria's basins attraction dynamical system. With we demonstrate increased convergence high risk, general-sum coordination games both via numerical analysis experiments.