Mixed Autonomy Traffic: A Reinforcement Learning Perspective

Cathy Wu
Microsoft Research AI, MIT

How might self-driving cars change urban mobility? Leveraging deep reinforcement learning and control theory, the talk begins to address this complex societal question by studying emergent behaviors in transportation dynamical systems, consisting of mixed automated and non-automated vehicles. By examining representative traffic phenomena, the work demonstrates that learned control laws can coordinate a small fraction of automated vehicles (5-10%) to eliminate traffic congestion and improve overall speeds or throughputs by 30-140%, and even achieve near-optimal performance in certain settings. The talk discusses advances in policy gradient methods and transfer learning motivated by mixed autonomy traffic. These results have implications for the near-term impact of self-driving cars, public health & safety, and urban sustainability.

Presentation (PDF File)

Back to Workshop II: Safe Operation of Connected and Autonomous Vehicle Fleets