Adaptive Punishment (APC) boosts AI cooperation in multi-agent games

Mixed-motive scenarios—where self-interested agents often defect for immediate rewards—are ubiquitous in multi-agent systems. Traditional peer punishment can deter defection, but as costly second-order altruism, it often undermines the punisher's own long-term gains. To solve this, researchers introduce Adaptive Punishment for Cooperation (APC), a distributed method that determines punishment intensity based on both a dynamic punishment probability and the severity of defection. This dynamic probability substantially reduces costly and ineffective punishment while promoting cooperation. APC includes a defection awareness module whose learning is guided by game reward, enabling accurate assessment of defection severity.

Empirically, APC demonstrates strong performance in iterated public goods game and significantly outperforms existing baselines across sequential social dilemmas. The method learns rational and effective punishment policies that foster cooperation by strategically deterring defection. Theoretical analysis supports the approach, showing it balances punishment cost and efficacy. This work has implications for designing cooperative AI systems in domains like autonomous vehicles, resource allocation, and economic simulations where agents must navigate mixed-motive interactions.