Graph Neural Network Controls Traffic Signals on Diverse Networks.

A new traffic-signal control interface has been developed, leveraging a graph neural network (GNN) to manage traffic flow. This system assigns scores to individual traffic movements, which each junction then converts into legal signal phases using a deterministic incidence matrix. The design separates the learned network from phase definitions and signal timing, making the graph size and junction-specific actions independent of the learned parameters. The GNN incorporates directed corridor nodes for traffic context and movement nodes representing paths through junctions. Experiments using Proximal Policy Optimization (PPO) evaluated the interface on unseen synthetic grid geometries, altered signal coverage, and five diverse city graphs. The policies maintained performance across new grid geometries but showed sensitivity to shifts in signal coverage. Notably, a single trained city-policy instance could operate across all five heterogeneous city graphs, albeit with varied outcomes, demonstrating the feasibility of this graph-based approach for complex road networks.