Toward Socially Friendly Autonomous Driving Using Multi-agent Deep Reinforcement Learning

We develop a novel multi-agent driving simulation framework (SFDPO) so that socially friendly driving behaviors can be acquired by agents through multi-agent reinforcement learning. We model personal and social driving behaviors in the driver model to reflect human driving goals and preferences. We make a game-theoretic assumption on fair compromised solution concepts to find an equilibrium solution under conflicts in complex interactive scenarios. A meta-policy optimization method is adopted to leverage personal and social driving behaviors in terms of personalized loss and socialized loss to achieve a balanced Pareto optimal solution between the socially friendly and personal preference driving goals.