Dynamic Autonomous Driving Headway Optimization with Deep Reinforcement Learning for Freeway Congestion Management and Control

Our transportation system is at the precipice of a paradigm shift with the significant and growing efforts that have been dedicated to developing Connected and Automated Vehicular Systems (CAVS) over the past few years. CAVS are expected to dramatically alter the capabilities of individual vehicles and the transportation infrastructure in smart cities of the near future. At present, some CAVS, in the name of comfort and safety, maintain long headways between vehicles which ultimately reduces road capacities in contrast to what we aspire for. As such, on one hand, the proliferation of CAVS can be detrimental to road capacities and traffic congestion, while on the other hand, if exploited correctly, they could create new opportunities for more efficient traffic control and management strategies through their automation and connectivity capabilities, leading to significant improvements in the capacities of our road infrastructure in addition to minimizing congestion and its negative externalities. This dissertation quantifies the expected impacts of CAVS on freeway driving and develops novel control strategies to maximize their benefits. Adaptive Cruise Control (ACC) systems are the core building block in future full autonomous driving. This dissertation examines two widely used ACC car following models and investigates the impact of the time headway parameter on traffic operation and performance on one of the busiest freeway corridors in Ontario, Canada. The dissertation further presents a regulator-based traffic control strategy which aims to adapt the time headway of ACC-equipped vehicles in real time according to the prevailing traffic conditions so that freeway performance is improved. It is then illustrated that, although shorter headways result in higher capacity, flow break down can still occur if traffic densities at bottlenecks are allowed to exceed the critical density which necessitates dynamic traffic control near bottlenecks to avoid bottleneck activation and capacity loss. Consequently, an adaptive Deep Reinforcement Learning (DRL) headway controller that uses ACC headways to optimize traffic flow and minimize delay based on real-time traffic information is proposed. The controller learns to dynamically assign an optimal headway value every control cycle for each controlled section on a freeway such that system delay is minimized, and flow maximized. This is first performed on a single-bottleneck network with a single global RL agent configuration to establish proper understanding of the approach and its impact, then on a multi-bottleneck network with a global RL agent and two different multiagent RL configurations. The proposed DRL headway control strategy was found to improve traffic, enhance the network throughput, and reduce the system delay by up to 57% and 46% in the single-bottleneck and the multi-bottleneck networks, respectively, compared to the examined nondynamic headways.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.