中央研究院 Logo

Social rules and conventions can emerge from interactions among individuals through gaming, mutual agreement, or state power. While the history, diversity, and justice of complex human rules and conventions have been extensively studied in the social sciences, how social norms arise in social animals—such as ants, whales, and elephants—remains elusive. Recent advances in artificial intelligence offer a new platform to revisit this question. Here, we use multi-agent reinforcement learning (MARL) as a computational laboratory to investigate how multiple agents reach mutual agreement on traffic rules. We use DQAttention and Social Residual Q-Network (SRQN) to train individual agents. Multiple such agents capable of navigation are then placed on the same map for further training. We consider various traffic scenarios, including two-lane roads, cross intersections, and roundabouts. We find that these agents can develop a multiverse of traffic norms, such as ranked passing priority, keep-left/right rules, and the first-come-first-go stop sign rule. By tuning the reward function or choosing whether to share parameters among agents, the agents can acquire different "characters," including aggressive or cooperative tendencies and left- or right-side keeping. Taken together, our work extends the study of collective behaviors in active matter to collective social behaviors and sheds light on how AI agents might develop their own social rules.