Traffic flow management of autonomous vehicles using deep reinforcement learning and smart rerouting

Abstract

Autonomous Vehicles (AVs) promise to disrupt the traditional systems of transportation. An autonomous driving environment requires an uninterrupted, continuous stream of data and information based on complex traffic data sets and predictive measurements to make critical and real-time decisions in uncertain situations. Such an environment fosters a self-organizing system where vehicles must be seamlessly connected and various other services and intelligent decisions to manage traffic flow must be executed in an emergent manner. To proceed towards this vision, in this paper, we develop a traffic flow management model which is based on a novel two-phase approach for AVs to optimize traffic during congestion periods. In the first phase of our approach, we build an adaptive traffic signal control, using Deep Reinforcement Learning (DRL) to optimize traffic flow on road intersections during the periods when traffic is congested. In the second phase, we implement a Smart Re-routing (SR) technique for the traffic approaching intersections. Re-routing is used to carry out load-balancing of traffic to alternate paths to avoid congested road intersections. The experimental evaluation of the proposed approach is validated using simulations that demonstrate up to 31% improved performance efficiency compared to the traditional settings using pre-timed signals and without re-routing. The two-phase approach improves the overall traffic flow while reducing delays and minimizing long traffic queues' lengths. This approach is useful for making infrastructure intelligent enough to handle traffic congestion and balance traffic flow efficiently.