Icon Legend

This session is not in your schedule.

This session is in your schedule. Click again to remove it.

96 Views

ViewAttendees 1

Abstract for presentation (Poster or Podium) with a Paper in the Conference Proceedings

Street & Highway Operations

Addressing Urban Traffic Congestion: A Hybrid DQN-Autoencoder Model with HyperOPT Tuning

Sunday, June 16, 2024

1:15 PM - 2:30 PM ET

Room: Galleria

Co-Author(s)

AB

Anurag Balakrishnan, n/a

Graduate Student
California State University, Long Beach
Long Beach, California, United States
SP

Satyam Pathak, n/a

Graduate Student
California State University, Long Beach
Long Beach, California, United States
PH

Pedro Herrera (he/him/his)

Undergraduate Student
California State University, Long Beach
Long Beach, CA, United States

Primary Author(s)

Tairan Liu, PhD (he/him/his)

Assistant Professor
California State University, Long Beach
Long Beach, CA, United States

Submitter(s)

TL

Tairan Liu, Ph.D.

California State University, Long Beach
Long Beach, California, United States

Abstract: In this study, we propose a novel method for managing traffic lights by integrating Deep Q-Networks (DQN) with Auto-encoders. The aim is to enhance traffic fluidity and mitigate congestion in simulated environments. To achieve this, we incorporate the average vehicle velocity as a key metric within the system's observation space. Additionally, we employ optimization techniques such as HyperOPT and leverage the data compression features of Auto-encoders to improve decision-making quality. Our approach was evaluated on a two-way, single-intersection network, subjecting to a vehicle flow rate varying between 100 and 600 vehicles per hour. We compared our hybrid DQN-Auto-encoder method against a DQN-only baseline, using average waiting time as the evaluation metric. The results indicate that our model markedly outperformed the baseline, reducing the average waiting time to approximately 200 seconds, as opposed to roughly 400 seconds observed in the baseline, while a traditional non-algorithmic approach yielded an average waiting time of nearly 1100 seconds.

Learning Objectives:

Attendees can expect to learn the following from this session:

demonstrate a significant reduction in waiting time in comparison to the current traffic intersections
integrate DQN, Auto-encoders, and HyperOPT to the traffic light control of complex traffic networks
alleviate urban traffic congestion challenges