Signal controllers learn whether a jam is a burst or a cut road: waiting-time recovery falls from 329 to 245 seconds under random blockages
Synopsis
Network signal control no longer compresses the whole road network into one shared latent vector. Policies trained under nominal flow often cannot tell a brief fluctuation from a sustained capacity loss when blockages, corridor failures or rain arrive. Here each intersection re-weights its neighbours from their joint state, then reads a history window to judge whether a change persists, cutting waiting-time recovery from 329 to 245 seconds under random blockages and by 82 seconds under a central closure.
Interpretation
Signal control now writes the perturbation into each intersection's own representation, instead of letting a coordination scheme fixed at training time absorb it. Earlier graph-attention methods produce a neighbour ranking shared across query nodes once trained, so the coordination pattern optimal under nominal flow carries unchanged into the disrupted regime; another line compresses the network into a single latent vector, leaving an agent unable to tell which part of its surroundings changed or for how long. On a controlled synthetic 4x4 grid, waiting-time recovery under random blockages is 245 seconds against MacLight's 329 seconds; under a central closure, recovery is 82 seconds faster for waiting time, 95 seconds for queue length and 81 seconds for average speed.
Each intersection looks at its neighbours first and then back over its own history, so the two steps jointly answer whom to cooperate with and which changes deserve a response. The spatial block scores each neighbour from the joint state of the pair using GATv2, so a capacity reduction reshapes the functional neighbourhood without changing the physical graph; the temporal block uses a Transformer encoder over recent spatial states that are already contextualized, separating short-lived noise from a lasting capacity loss. Observations cover the active phase, a minimum-green indicator, lane density and queue saturation, actions are eight feasible phases, and the policy is learned with PPO under centralized training and decentralized execution; evaluation runs in SUMO 1.19 on the synthetic 4x4 grid and the Hangzhou 4x4 network, with five random seeds per scenario.
Disruptions are turned into a reproducible stress test rather than simply heavier traffic. Conventional evaluation compares efficiency under fixed demand files and rarely isolates structural failures as separate scenarios. The suite includes random blockages that resample eight edges every 300 seconds, west-east and south-north corridor failures, a central square closure, and a rainy scenario that scales desired speed by 0.80 and raises the minimum gap from 2.5 to 3.0 metres.
Perspective
The controller is evaluated on a controlled synthetic 4x4 grid and the Hangzhou 4x4 network, with perturbation and recovery analysis restricted to the synthetic grid so that blockage location, duration and severity stay reproducible. Observations are limited to phase, minimum-green indicator, lane density and queue saturation; actions are eight feasible phases; execution uses only local observations and directly connected neighbours. It suits researchers and engineers working on robustness evaluation for signal control.
Under nominal conditions the waiting-time advantage is small relative to the observed variability, and the authors present it as competitive performance rather than statistical superiority. In the south-north corridor failure the method records a higher degradation rate than MacLight, which the authors attribute to a baseline effect because its nominal waiting time is lower, while its absolute waiting time stays lower. The rainy scenario only changes desired speed and minimum gap, which the authors describe as a literature-informed sensitivity setting rather than evidence of robustness under real rain.