LIM Center, Aleje Jerozolimskie 65/79, 00-697 Warsaw, Poland
+48 (22) 364 58 00

Advancing Congestion Modeling: Deep Reinforcement Learning in Smart Transportation

Advancing Congestion Modeling: Deep Reinforcement Learning in Smart Transportation

Advancing Congestion Modeling: Deep Reinforcement Learning in Smart Transportation

In a recent webinar hosted by the VinFuture Foundation, top professionals from the fields of smart transportation and technology discussed the potential of deep reinforcement learning in easing traffic and enhancing transport system management. The webinar brought together experts from both academic institutions and Vietnamese businesses to share knowledge and explore collaborations.

Dr. Padmanabhan Anandan, a renowned expert in computer vision and artificial intelligence, chaired the webinar. He highlighted three key factors that have made advancements in smart transportation and mobility possible: the ubiquity of sensors, intense communication between vehicles and the system through the internet, and the ability to control various aspects of driving through software.

One of the promising developments discussed in the webinar was the MegaVander Test, which focuses on mixed autonomy in traffic. This condition involves the coexistence of controlled automated vehicles (CAVs) and human-driven vehicles on the same road. The goal is to leverage technology to reduce congestion and increase safety.

Prof. Alexandre Bayen from the University of California, Berkeley, presented the latest advances in transportation management technology. He emphasized that a major contributor to traffic congestion is the stop-and-go wave phenomenon caused by human-driven vehicles. Controlled automated vehicles equipped with GSM or 5G chips can improve traffic flow through coordinated communication among themselves. Initial findings suggest a potential 24% reduction in total energy consumption.

The control system’s algorithm ensures a safe distance between vehicles and manages speed on highways. To address the lack of data and loss of communication, the team has implemented two layers of controllers: an expert controller that relies on human calculations, and an imitation learning controller that operates even in the absence of connectivity.

Machine learning plays a crucial role in transforming transportation research. Prof. Bayen discussed two approaches: imitation learning and deep reinforcement learning. Imitation learning enables the mega-controller system to operate based on demonstrated behaviors, while deep reinforcement learning allows CAVs to evaluate situations and make decisions based on a predefined algorithm.

The MegaVander Test has been conducted on real highways, and preliminary findings show its effectiveness in reducing congestion and saving energy. The team is currently analyzing data to assess its impact further.

Overall, deep reinforcement learning has the potential to revolutionize congestion modeling and smart transportation, leading to more efficient and sustainable urban mobility solutions.

– InnovaTalk webinar on Smart Transportation and Mobility Solutions for Urban Areas
– VinFuture Foundation
– University of California, Berkeley’s transportation research team