Abstract

Rapid urbanization, increasing travel demand, and limited road capacity have intensified congestion, delays, fuel consumption, and vehicular emissions in urban transportation systems. This study evaluated a Deep Q-Network (DQN)-based adaptive traffic signal controller in a simulated multi-intersection network representative of Tehran. A microscopic traffic network comprising 24 signalized intersections was developed using the Simulation of Urban Mobility (SUMO) platform. The controller uses a 48-dimensional state representation and selects one of eight discrete control actions. Its reward function incorporates queue length, waiting time, and traffic throughput, whereas fuel consumption and CO₂ emissions are evaluated independently as environmental indicators. The DQN was compared with fixed-time and semi-adaptive signal-control strategies using average delay, queue length, average speed, fuel consumption, and CO₂ emissions. Each strategy was evaluated using 30 independent simulation replications. Normality and homogeneity of variances were assessed using the Shapiro–Wilk and Levene tests, followed by one-way analysis of variance (ANOVA) and Tukey’s honestly significant difference (HSD) post hoc test. Under simulated conditions, DQN reduced average delay from 78.4 to 45.2 s and average queue length from 42 to 22 vehicles relative to fixed-time control, corresponding to reductions of 42.35% and 47.62%, respectively. Average speed increased from 21.5 to 31.4 km/h, representing a 46.05% improvement. Fuel consumption decreased from 512 to 417 L/h, whereas CO₂ emissions decreased from 1,280 to 1,042 kg/h, corresponding to reductions of 18.55% and 18.59%, respectively. Differences among the strategies were statistically significant for the principal indicators (p < 0.001). Sensitivity analysis across demand levels of 80%–120% of the baseline indicated a stable performance. Overall, the DQN improved the selected traffic and environmental indicators under simulated conditions. However, the findings are specific to the modeled network and simulation assumptions; field calibration, independent validation, and real-world testing are required before deployment.