Duration: 19:09

72 views

0

Collisions, crashes, and other incidents on road networks, if left un-mitigated, can potentially cause cascading failures that can affect large parts of the system. In this work, we investigate the efficacy of various Deep Reinforcement Learning (DRL) based AI algorithms in reducing traffic congestion on multi-lane freeways in such adverse scenarios.

Our goal is to design re-routing strategies for optimal and adaptive utilization of the freeway lanes and the arterials in proximity to reduce the resulting congestion, enable the emergency management, and improve the speeds. We pose the problem as a Markov Decision Process (MDP) and use DRL-based algorithms such as Deep Q-Network (DQN) and the Advantage Actor Critic (A2C) approaches to train an autonomous agent. For this purpose, we build a robust interface between the DRL libraries and the SUMO traffic simulator.

The test network is a 2.6-mile-long 4-lane freeway stretch in Shoreline, Washington, USA with two exits and associated arterial roads. The injected temporal traffic profiles are parameterized from real-world traffic counts data to generate training and deployment scenarios that include incidents. The speed and the count data estimated by the virtual sensing within SUMO is used to define the system state and the re-routing actions recommended by the DRL algorithms are actuated by the TraCI controller interface.

We study how re-routing the traffic via dynamic messaging signs can help alleviate congestion by leveraging multiple reward functions. Our study also illustrates the use of transfer learning on a simple example in which the agent is trained on regular congestion data and deployed on the scenarios involving accidents. Using the flow-speed-density based macroscopic fundamental diagram, we explain how the learning accomplished by the DRL agent and the trade-offs discovered automatically, are consistent with the first principles of traffic theory.