Automated Planning for Urban Traffic Management

Thomas Leo McCluskey, Mauro Vallati, Santiago FrancoIJCAI 2017 · DOI

TL;DR

This paper presents SimplyfAI, a system that uses automated AI planning to manage urban traffic, specifically focusing on optimizing traffic light timings to mitigate air pollution and reduce congestion. By leveraging PDDL+ for a rich representation of traffic flows and interventions, the system can generate real-time, effective strategies for complex traffic scenarios, like road closures, demonstrating a significant step towards more intelligent urban mobility.

What Problem Does This Paper Address?

The increasing urbanization worldwide is leading to significant challenges in managing urban road networks and their associated traffic emissions. Cities struggle to meet air quality limits due to the limited effectiveness of current traffic management devices. While AI has been applied to transportation, there's a need for more sophisticated planning approaches that can handle exceptional circumstances (like road closures) and proactively manage traffic to prevent pollution problems before they occur. Existing methods often rely on microscopic simulations which have scalability issues, or simpler planning representations that may not capture the nuances of traffic flow.

Methods

The SimplyfAI system employs an AI planning approach, specifically using the PDDL+ language to model urban traffic. PDDL+ allows for a more granular representation of traffic by describing road link densities (e.g., low, medium, high) and continuous variables, which is crucial for capturing traffic flow dynamics. The system's architecture involves:

  • Domain Engineering: Experts define PDDL+ models of transport interventions, such as traffic light changes and flow rates on road links, along with static information about the road network.
  • Real-time Planning: When triggered by a goal (e.g., reducing congestion or emissions), the system uses the UPMurphi planner to generate plans. This involves reformulating the PDDL+ model for efficient reasoning and exploiting domain-specific heuristics.
  • Strategy Generation and Refinement: The planner generates "strategies" – temporal schemes of traffic light interventions. A heuristic approach is then used to refine these strategies by identifying and removing traffic lights that have minimal impact on the goal, thereby optimizing the overall strategy and minimizing disruption.

Datasets

Not discussed in the paper. The paper mentions using historical records for expected vehicle flow and a simulated world for evaluation, but specific datasets are not detailed.

Key Insights and Ideas

  • PDDL+ for Rich Traffic Representation: Using PDDL+ allows for a more accurate and granular representation of traffic flows and interventions compared to simpler planning formalisms, bridging the gap between planning and traffic simulation.
  • Proactive Traffic Management: The system aims to prevent air pollution problems before they occur, moving beyond reactive traffic control.
  • Handling Exceptional Events: The planning approach is particularly effective in generating real-time strategies for unexpected situations like road closures, ensuring efficient re-routing.
  • Heuristic Refinement of Strategies: Optimizing generated plans by removing non-essential traffic light controls is a key insight for improving the efficiency and minimizing the impact of the overall traffic management strategy.
  • Scalability through Macroscopic Modeling: By modeling traffic at a macroscopic level (density rather than individual vehicles), the system overcomes the scalability limitations of microscopic simulation approaches.

Key Results

The paper demonstrates that the SimplyfAI planning system can generate effective strategies for dealing with exceptional situations and reducing emissions. Specifically, the system can:

  • Generate real-time plans (in seconds) for traffic light changes in response to events like road closures.
  • Efficiently re-route vehicles by adjusting traffic light phases.
  • Produce strategies that are comparable to, and can be visually demonstrated against, manually optimized green light phases designed by traffic experts using simulators like AIMSUN.
  • The approach is shown to be effective in managing traffic at a regional level.

Limitations and Future Work

  • Scalability of Planning: While the macroscopic approach improves scalability, the paper acknowledges that the complexity of the PDDL+ model can still pose challenges for the planner. The reformulation process aims to mitigate this.
  • Centralized Control Assumption: The paper notes that centralized planning to control individual cars assumes knowledge of every vehicle's position and the ability to communicate with them, which is a significant practical hurdle. The SimplyfAI approach focuses on traffic light control, which is a more feasible intervention.
  • Future Work: The authors suggest extending the system to incorporate other interventions, such as variable speed limits, and further exploring more subtle goals related to limiting vehicle emissions directly. They also highlight the potential for integrating this planning system with traffic simulators for more comprehensive validation.

Summary updated

BibTeX
@inproceedings{mccluskey2017automated,
  author = {Thomas Leo McCluskey and Mauro Vallati and Santiago Franco},
  title = {{Automated Planning for Urban Traffic Management}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2017},
  doi = {10.24963/ijcai.2017/776},
}
Figure