Urban road networks underpin commuting, emergency response, freight movement, and everyday economic activity. Yet a crash, flood, extreme-weather event, or other sudden loss of road capacity can trigger effects far beyond the initially affected link. Queues spread, travelers change routes and departure times, bottlenecks migrate, and the network may undergo an extended adjustment process before reaching a new operating state. The challenge for traffic managers is therefore not only to limit the immediate loss of performance, but also to accelerate recovery.
Infrastructure expansion and added redundancy can strengthen a network, but they are costly and slow to deploy. Traffic signals offer a more flexible operational lever. Conventional fixed-time, capacity-based, and equi-saturation strategies, however, are not designed to learn from the coupled evolution of congestion and traveler behavior during a disruption. A study published in Engineering addresses this gap by asking whether signal timing can be trained specifically to protect network resilience rather than merely improve traffic conditions at one moment in time.
The research team developed a deep Q-network (DQN) framework in which traffic signals act as learning agents. The agents observe link flow and link capacity and adjust red-green time splits. The controller is coupled with a within-day/day-to-day dynamic traffic assignment model: the within-day component represents traffic propagation, queuing, and congestion dissipation, while the day-to-day component captures how travelers revise routes and departure times in response to their previous experience. This integration allows the controller to learn within an environment where traffic states, signal decisions, and traveler choices continually influence one another.
One of the key innovation is the resilience-oriented reward. The relative area index (RAI) measures the cumulative gap between disrupted network performance and its normal baseline. A smaller RAI means that the network experiences less overall performance loss and/or recovers more quickly. By placing RAI inside the reward function, the learning process targets both the depth and the duration of disruption-induced degradation.
The method was evaluated on the Sioux Falls and Anaheim benchmark networks. The experiments represented minor, moderate, and severe disruptions by reducing selected link capacities by 25%, 50%, and 75%, respectively. Three-, four-, five-, and six-leg intersection scenarios were examined, and the DQN-based controller was compared with equi-saturation, capacity-based, and fixed-time control. All convergence results are reported in modelled day-to-day iterations. Each iteration represents one cycle of traveler learning and network-state updating, so it is a theoretical measure of adjustment and convergence-not a calendar day in a real city.
In Sioux Falls, the proposed controller produced the lowest RAI across all tested intersection types and disturbance levels. In the five-leg moderate-disruption scenario, the network reached equilibrium after 28 modelled behavioral-adjustment iterations with an RAI of 0.3547, which is 23.1%, 69.4%, and 71.0% lower than under equi-saturation, capacity-based, and fixed-time control, respectively. In the three-leg severe-disruption scenario, it reached a new equilibrium after 38 iterations, compared with 49, 70, and 59 iterations for the same three baselines. The corresponding RAI reductions were 49.6%, 69.2%, and 65.6%, while the number of iterations required to reach equilibrium fell by 22.4%-45.7%.
The benefit remained substantial in the five-leg severe-disruption scenario: the proposed controller reached equilibrium after 32 modelled adjustment iterations with an RAI of 4.7097, reducing RAI by 23.8%, 32.1%, and 67.0% relative to the three conventional strategies and requiring 13-18 fewer iterations than the baselines. Link-level analysis showed how these network-wide gains emerged. For example, the travel cost on link 26 at disrupted intersection 9 was 44.5% lower than under capacity-based control, while the cost on link 29 was reduced by as much as 63.7% relative to equi-saturation control. These changes indicate that adaptive timing can curb cost accumulation at critical bottlenecks and limit the spread of localized congestion.
An ablation experiment confirmed that the reward design mattered. In the representative five-leg severe-disruption case, the RAI-based reward achieved an RAI of 4.7097 and stability after 32 iterations. A reward based only on total travel cost produced an RAI of 5.5535 and stability after 38 iterations, while a reward focused only on convergence time produced an RAI of 5.7261 and stability after 33 iterations. The RAI reward therefore offered the strongest balance between limiting cumulative loss and restoring stable operation.
The larger Anaheim network provided a further test of scalability. Under a severe disruption at six-leg hub intersections, the DQN-based strategy reached equilibrium after 51 modelled adjustment iterations with an RAI of 1.8660. This was 44.6%, 59.4%, and 51.7% lower than the values obtained with equi-saturation, capacity-based, and fixed-time control, respectively.
The findings reposition traffic signal control as a tool for disruption recovery, not only routine congestion management. Because the framework explicitly links signal actions, network dynamics, traveler adaptation, and a resilience metric, it could support contingency planning, digital-twin testing, and decision support for emergency traffic organization. The reported evidence is simulation-based, however. Before field deployment, the controller would need calibration with local detector and incident data, compliance with signal-safety and operational constraints, and validation under real-world disturbances with human oversight.
By training signals to minimize the accumulated consequences of disruption, the study offers a measurable and transferable route toward more robust and faster-recovering urban transport systems.
The paper “Traffic Signal Control with Deep Reinforcement Learning Toward Enhanced Resilience of Urban Road Networks,” is authored by Qiannian Xiang, Zaoli Yang, Haibo Chen, Washington Ochieng, Daoping Wang, Chi Xie, and Wen-Long Shang. Full text of the open access paper: https://doi.org/10.1016/j.eng.2026.05.016. For more information about Engineering, visit the website at https://www.sciencedirect.com/journal/engineering.