Model shows that deep learning can make traffic signals up to 75 percent more efficient.
The analysis, which takes into account an area with multiple intersections, consistently shows superior performance.
Made Public Date
01/27/2020
Identifier
2020-B01435
TwitterLinkedInFacebook

Cooperative Traffic Signal Control with Traffic Flow Prediction in Multi-Intersection

Summary Information

Traffic signal control, which seeks to optimize the behavior of signalized intersections, is a focus of much research and analysis. By making intersections as efficient as possible, transportation engineers may minimize the impacts of congestion on dense areas. A paper by researchers from Gachon University in South Korea proposes a cooperative traffic signal control with a traffic flow prediction (TFP-CTSC) for multi-intersection spaces. Traffic flow prediction models simulate future traffic states from observed variables affecting current traffic conditions. The cooperative signals are designed to share information across intersections, allowing for behavior that is globally, rather than locally, efficient.

The proposed algorithm uses deep Q-network (DQN) reinforcement learning, which uses a neural network to select optimal actions based on the existing state of the environment and building off of previous saved experiences. The authors adapted DQN to create a TFP-CTSC algorithm that seeks to minimize the time spent waiting by vehicles within the network. Additionally, the TFP-CTSC algorithm considers the state of other nearby intersections to improve its predictions, which previously had not been incorporated in traffic signal learning algorithms.

To test the proposed model, the authors conducted an experiment comparing their TFP-CTSC algorithm with previously developed DQN and Q-learning algorithms, which did not account for other intersections. The authors then implemented the algorithms in a series of simulations to compare their performance.

FINDINGS

The experiment indicated that the TFP-CTSC algorithm was both more efficient and more consistent than the other algorithms tested.

  • Early on in the learning process, the Q-learning algorithm reached a peak average waiting time of more than 150 seconds and the DQN algorithm reached a peak waiting time of over 100 seconds, while the TFP-CTSC algorithm peaked at a maximum average waiting time under 50 seconds. All three algorithms became more effective as the simulations progressed.
  • For mature models, the average waiting time ranged between 40 and 80 seconds for the Q-learning algorithm and between approximately 15 and 30 seconds for the DQN algorithm, but ranged between 10 and 20 seconds for the TFP-CTSC algorithm.
  • The Q-learning and DQN algorithms had similar average queue lengths, varying between roughly 60 and 30 vehicles, while the TFP-CTSC algorithm maintained a consistent average queue length of only 10 vehicles.
  • The average overall delay time for the Q-learning algorithm was slightly greater than 35 seconds; for DQN, it was approximately 30 seconds; for the TFP-CTSC algorithm, it was approximately 20 seconds.
Goal Areas
Deployment Locations