Whole-network one-way switching: about two thirds of roads flip to the peak direction at rush hour, cutting daily delay by roughly 80%

Every road's lane direction and directional speed limit can now be re-decided together at each decision step, instead of following fixed directions or corridor-level reversible-lane schedules. Earlier dynamic lane reversal was mostly solved as integer or bi-level programs and dynamic traffic assignment over given or forecast demand, and reinforcement-learning lane allocation kept at least one lane in each direction of every road. On a grid with commuter rush hours, the learned policy makes about two thirds of roads one-way in the peak direction and cuts daily delay by roughly 80% against the fixed initial configuration and by roughly 70% against a tuned two-way load balancer.

Lane proposals do not take effect directly; they pass a check that accepts a proposal closing a direction only if the network stays strongly connected. Giving all lanes of a road to one direction turns it into a one-way road that can cut junctions off, and Robbins' theorem covers only a single orientation that never changes. Closing proposals are considered one at a time in random order; a rejected one is relaxed to keep one lane open on multi-lane roads or reverted on single-lane roads, and lane clearing before each reversal is modeled explicitly, so strong connectivity holds at every time step, including during clearing.

All road controllers share one transformer whose attention runs jointly over all road-time tokens and is biased by the endpoint relations between roads. Earlier spatio-temporal attention models usually attend over space and time in separate steps, and graph transformers inject structure through shortest-path distance and node degree. The model is pretrained on action-conditioned traffic forecasting and fine-tuned with per-road proximal policy optimization under a mixed reward combining the road, its neighbors and the network; pretraining roughly halves the updates needed to surpass the fixed configuration.

Ablations show that removing the relational bias or the spatial embedding multiplies the final policy's delay about fivefold, while the other three components matter less and their earlier evaluations disagree with pretraining. The other three ablations, collapsed endpoint relations, an undirected spatial embedding and no temporal embedding, end within about a few percent of the full model in return but incur one to two times its delay. Only for the temporal embedding do the earlier evaluations and the pretraining error agree; the collapsed-relations and undirected-embedding variants led the full model at most earlier evaluations.

The demonstrated setting is a synthetic grid network with morning and evening commutes between two opposite corners, all vehicles homogeneous connected and automated vehicles that comply fully and route with knowledge of current traffic, simulated mesoscopically without signals, turning conflicts or capacity drop at junctions.

The learned speed limits add delay at test time, and holding them nominal removes more than half of the remaining delay; the authors note the simulator lacks the congestion mechanisms on which variable speed limits act, so whether learned speed limits can contribute to joint control remains open. Each variant is trained once, so return differences of about a few percent between variants cannot be firmly attributed to the architecture; the collapsed-relations and undirected-embedding variants incur higher final delay yet led the full model at most earlier evaluations.