2664. Lane-Level Traffic Flow Estimation with Federated Learning for Urban Vehicle Routing
Invited abstract in session MD-15: Vehicle Routing problems, stream Combinatorial Optimization.
Monday, 14:30-16:00Room: Esther Simpson 1.08
Authors (first author is the speaker)
| 1. | Chansoo Kim
|
| AI, Information and Reasoning (AIR) Lab., Computational Science Centre, Korea Institute of Science and Technology (KIST) & University of Science and Technology (UST) | |
| 2. | Sangchul Lee
|
| AI, Information and Reasoning Lab., Computational Science Centre, Korea Institute of Science and Technology |
Abstract
We propose an integrated approach for lane-level traffic flow estimation for application to the Vehicle Routing Optimization (VRO) problem. Traditional traffic models, based on average flows, fail to capture nonlinear dynamics and bottlenecks at intersections, ramps, and merging areas. To address this, we construct a graph-based lane-level model incorporating network flow constraints.
To improve flow estimation, we introduce a deep learning-based auxiliary module, complemented by a stochastic agent-based traffic model to analyze complex lane interactions. The predicted flows serve as key inputs to a Mixed-Integer Programming (MIP) model formulated to solve the VRO problem over a lane-level network. By providing lane-level data, the estimation module enhances the accuracy of decisions within the optimization framework. This model helps determine vehicle routing and allocation under cost functions reflecting actual traffic conditions.
We also apply decentralized federated learning to integrate locally trained models across distributed networks. Preliminary simulations on a synthetic complex urban traffic network show approximately 12.3% congestion reduction compared to average flow-based approaches. Sensitivity analyses suggest the scalability under diverse traffic patterns. This study highlights potential of combining AI-assisted estimation with flow optimization to address global issues such as smart cities and autonomous driving, with scalability to multimodal networks.
Keywords
- Simulation
- Vehicle Routing
- Artificial Intelligence
Status: accepted
Back to the list of papers