Agent-based Real-time Signal Coordination in Congested Networks

This study is the continuation of a previous NEXTRANS study on agent-based reinforcement learning methods for signal coordination in congested networks. In the previous study, the formulation of a real-time agent-based traffic signal control in oversaturated networks was described and exemplified through a case study. The agent-based control was implemented using two different reinforcement learning algorithms: Q-learning and approximate dynamic programming. Also, the performance of the network was evaluated using the max-plus algorithm to provide explicit coordination between the agents. The RL algorithms and max-plus showed satisfactory performance and were able to efficiently process traffic, reducing the frequency of queue spillbacks and preventing gridlocks. This study extends the previous implementations and describes the use of explicit coordinating mechanisms with Q-learning, mainly through a modified max-plus version developed throughout this research project. A traffic network similar to that in the previous study is used to compare the results without explicit coordination, with the standard max-plus and the enhanced coordination. Results indicate that the enhanced coordination has the potential to further improve signal operation mainly by reducing the number of stops per vehicle, while maintaining an efficient vehicle processing rate. In addition, two more topics were explored and are presented in this report: the use of a function approximation to reduce memory requirements from large lookup tables and speed up convergence by means of generalization, and the effects of imperfect information received by the agents or faulty detectors. The case studies analyzed in this report are focused on oversaturation and thus, on managing traffic efficiently while preventing queue spillbacks and gridlocks. In this sense the applications described here do not only consider closely-spaced intersections in a grid-like network, but also high demands in all directions, resulting in scenarios where signal control is not straightforward. For this reason, it is expected that the findings in this report are also applicable to less challenging scenarios with similar configurations.

Language

  • English

Media Info

  • Media Type: Digital/other
  • Edition: Final Report
  • Features: Figures; References; Tables;
  • Pagination: 60p

Subject/Index Terms

Filing Info

  • Accession Number: 01537968
  • Record Type: Publication
  • Report/Paper Numbers: 0102IY04
  • Contract Numbers: DTRT07-G-005
  • Files: UTC, TRIS, RITA, ATRI, USDOT
  • Created Date: Sep 9 2014 9:14AM