HLSR: Hybrid Live Forecast Selective Dynamic Vehicl... | AI Research

HLSR (Hybrid Live–Forecast Selective Dynamic Vehicle Rerouting) is a framework designed to reduce urban traffic congestion by selectively rerouting vehicles based on a combination of real-time traffic data and short-term future predictions. By focusing only on vehicles affected by congestion rather than replanning the entire network, the system aims to improve travel times while minimizing unnecessary route changes.

Addressing Congestion with Selective Rerouting

Urban traffic management often struggles with two extremes: either ignoring vehicles already on the road or attempting to replan every vehicle in the network, which is computationally expensive and potentially disruptive. HLSR addresses this by using a three-phase approach: detecting congestion using dual thresholds for occupancy and velocity, selecting only the relevant vehicles for intervention, and allocating alternative routes. This selective strategy prevents the system from over-rerouting distant vehicles or failing to react to traffic approaching a bottleneck.

How the Hybrid System Works

The core of HLSR is its hybrid costing mechanism, which blends current live traffic speeds with short-horizon forecasts. Instead of relying solely on present conditions, the system uses a spatio-temporal forecaster (LSTAN_GERPE) to predict how traffic will evolve over the next few decision periods.
When calculating the best route for a vehicle, the system applies a horizon-dependent weight:

• Near-term segments: The system prioritizes live sensor data, as current observations are the most reliable.

• Future segments: The system relies more heavily on the model’s predictions to anticipate conditions the vehicle will encounter later in its journey.
  This process is further refined by a driver-tailored module that adjusts these network-level predictions to account for individual driving behaviors, ensuring that route allocations are more accurate for specific vehicles.

Performance and Results

In tests using a reproduced Tainan SUMO traffic scenario, HLSR demonstrated significant improvements in travel efficiency. At a volume of 8,000 vehicles, HLSR achieved a mean travel time of 380.6 seconds. This outperformed selective live-only controls (438.1 seconds) and scoped live Dijkstra methods (391.6 seconds). The framework also remained competitive with network-wide live travel-time Dijkstra (408.1 seconds) while requiring intervention on significantly fewer vehicles. These performance gains persisted at higher traffic volumes of 16,000 and 20,000 vehicles.

Considerations for Implementation

While HLSR improves upon existing methods, it is not a stand-alone end-to-end rerouting policy. The forecasting module is designed as a supporting component that feeds into the broader rerouting loop. Additionally, the authors note that while the system uses a ranking-aware fine-tuning process to ensure that predicted speeds preserve the correct order of candidate routes, the effectiveness of the system depends on the accuracy of the underlying traffic-light queueing models and the quality of the input data from roadside sensors.