Traffic signal coordination is usually modeled as a multi-agent problem: each intersection has its own agent, and agents learn to cooperate. This captures the decentralized nature of real traffic systems but creates coordination challenges — thousands of agents must learn compatible policies, which scales poorly and converges slowly.
A single-agent RL model for large-scale regional traffic control (arXiv:2503.09252) eliminates the coordination problem by treating the entire signal network as a single decision. One agent controls all intersections simultaneously, optimizing total travel time across the region rather than per-intersection metrics.
The counterintuitive result from robustness evaluation (arXiv:2603.15283): RL agents trained on diverse traffic patterns outperform traditional actuated signal control even under unseen demand scenarios. The model generalizes to traffic patterns it was never trained on — not because it learned universal traffic rules, but because diverse training creates robustness to distributional shift. The variety of training experience matters more than the fidelity of any single training scenario.
This challenges the dominant approach in traffic RL, which optimizes for specific traffic patterns (morning rush, evening rush, weekend). Training on the “right” pattern produces a brittle agent that fails on the “wrong” pattern. Training on all patterns produces a robust agent that performs adequately everywhere.
The engineering lesson: for systems operating in non-stationary environments, the single robust policy often outperforms the collection of specialized policies, because the real world spends most of its time between the scenarios you planned for. The edges of your training distribution are closer to reality than the centers are.
HALO achieves 6.8% lower average travel time across 2,668 intersections — suggesting the approach scales to real cities.