Incorporating Queueing Dynamics into Schedule-Driven Traffic Control

Hsu-Chieh Hu, Allen M. Hawkes, Stephen F. SmithIJCAI 2021 · DOI

TL;DR

This paper introduces a novel approach to real-time traffic signal control by incorporating dynamic queueing behavior into predictive models. By accounting for how queues evolve as different signal timing plans are considered, the system achieves a significant reduction in average vehicle wait times, particularly in heavy traffic conditions.

What Problem Does This Paper Address?

Optimizing traffic flow in urban environments is a critical challenge, and real-time traffic signal control systems offer a promising solution. Schedule-driven traffic control, a leading approach, relies on predicting when vehicles will arrive at and pass through intersections to generate optimal signal timing plans. However, existing methods typically use a static vehicle arrival model. This simplification fails to capture a crucial dynamic: the intersection's queue length and the resulting delays change as the system explores different potential signal timing schedules. This oversight can lead to suboptimal predictions and, consequently, less efficient traffic management, especially when queues are substantial. The paper aims to bridge this gap by developing a more accurate predictive model that dynamically accounts for queueing dynamics.

Methods

The core innovation of this paper is the development of an online queue prediction algorithm that integrates queueing dynamics into the forward search process for signal timing schedules. Instead of assuming a static queue, the algorithm dynamically projects incremental queueing delays for each vehicle as it considers different future signal timing scenarios.

Key aspects of the method include:

  • Dynamic Queueing Delay Projection: For each potential signal timing schedule (search state), the algorithm calculates an incremental queueing delay for vehicles. This delay is based on the previous state, the projected free-flow arrival time, and a queue dissipation model.
  • Queue Dissipation Model: The paper utilizes a linear model, qdtotal(Nq) = ao + a1 * (Nq - 1), to estimate the total queueing delay for a vehicle joining an existing queue of Nq vehicles. ao and a1 are parameters that can be learned from data.
  • Intra-cluster and Inter-cluster Queueing Delay: The model distinguishes between delays experienced by vehicles within a cluster (intra-cluster, affecting departure time) and delays experienced by the first vehicle of a cluster due to preceding queued clusters (inter-cluster, affecting the permitted start time).
  • Integration into Objective Function: The projected queueing delay is added to the cumulative delay caused by signal operations, providing a more accurate objective for the planning process.
  • Algorithm 1 & 2: These algorithms detail the calculation of inter- and intra-cluster queueing delays and the update of state variables (phase duration, finish time, total delay) during the search process.
  • Algorithm 3: This algorithm describes how to predict the queue count at a given state by considering dissipated vehicles and newly arriving vehicles.

Datasets

The paper does not explicitly mention training datasets. For evaluation, they used:

  • Microscopic Traffic Simulation: A real-world road network model of the St. Albert neighborhood in Canada, consisting of 11 intersections.
  • Simulated Traffic Demand: Traffic demand was generated based on field data from St. Albert, Canada, and scaled to create low, medium, and high demand scenarios for both AM and PM rush hours.

Key Insights and Ideas

  • Static Models are Insufficient: The most significant insight is that static queue models are a bottleneck for advanced traffic control. Ignoring how queues evolve dynamically as potential futures are explored leads to suboptimal decisions.
  • Queue Dynamics are Predictable: The paper demonstrates that queueing dynamics, while complex, can be modeled and predicted with sufficient accuracy using relatively simple linear models and by tracking vehicle arrivals and departures within the planning horizon.
  • Improved Arrival Time Prediction: By incorporating queueing dynamics, the system achieves a more accurate prediction of vehicle arrival times, which is fundamental to the effectiveness of schedule-driven control.
  • Trade-offs in Traffic Control: The results highlight that simply optimizing for queue stability or using reactive backpressure methods is less effective than a predictive approach that accounts for future queue states.

Key Results

The proposed approach demonstrated significant improvements in traffic flow efficiency:

  • Average Wait Time Reduction: A 10-15% reduction in average wait times was observed in heavy traffic scenarios compared to a schedule-driven system using the static predictive model.
  • Delay and Stop Improvements: Compared to the benchmark (queue stability) and bi-directional extension methods, the proposed approach showed improvements of approximately 13%/19% in delay and 24%/23% in the number of stops for AM high traffic demand.
  • Backpressure Outperformance: The algorithm achieved an improvement of about 20% in delay and 48% in the number of stops compared to a backpressure adaptive control method.
  • Reduced Queue Counts: Figure 4 shows that the proposed approach resulted in shorter average queue counts at congested intersections compared to the benchmark and backpressure methods.
  • Prediction Accuracy: Table 2 indicates that the queue prediction model can achieve a prediction error of around 1 vehicle with a lookahead of up to 40 seconds.

Limitations and Future Work

  • Parameter Sensitivity: The accuracy of the queue dissipation model relies on the parameters ao and a1. While the paper suggests these can be learned, their sensitivity and optimal learning strategies are areas for further exploration.
  • Model Complexity: While the linear queue dissipation model is effective, the paper notes that more complex polynomial or non-linear forms could be explored for even greater accuracy, though this might increase computational complexity.
  • Real-World Deployment: The paper's evaluation is based on simulation. Further validation through real-world deployment and testing would be valuable.
  • Integration with Other Control Strategies: Future work could explore integrating this queueing dynamics model with other advanced traffic control strategies or connected/autonomous vehicle functionalities.

Summary updated

BibTeX
@inproceedings{hu2021incorporating,
  author = {Hsu-Chieh Hu and Allen M. Hawkes and Stephen F. Smith},
  title = {{Incorporating Queueing Dynamics into Schedule-Driven Traffic Control}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2021},
  doi = {10.24963/ijcai.2021/561},
}
Figure