Empirical Evaluation of Reinforcement Learning for Adaptive Traffic Signal Control in Nigerian Urban Intersections

Adaptive traffic signal control using reinforcement learning has demonstrated significant reductions in vehicle waiting time in simulated and international urban environments, but no empirical evaluation exists for Nigerian urban intersection characteristics, which differ from Western contexts in driver behaviour, heterogeneous vehicle mixes, and informal pedestrian crossing patterns. This study empirically evaluated deep reinforcement learning (Deep Q-Network and Proximal Policy Optimisation) for adaptive signal control at a realistic model of the Berger Roundabout in Abuja, developed using the SUMO traffic simulation environment. A traffic flow model was calibrated against 6 hours of video-based traffic count data collected at the intersection under three traffic density conditions. The DQN and PPO agents were trained for 500,000 environment steps using a reward function that penalised cumulative vehicle waiting time per phase transition. The trained DQN agent reduced mean vehicle waiting time by 29.4 percent and queue length by 23.1 percent compared to the calibrated fixed-time signal baseline. PPO performed comparably to DQN after training, with slightly faster convergence. Both RL controllers outperformed an adaptive Webster-based algorithm by 14.2 and 11.7 percentage points respectively in waiting time reduction. Results remained consistent under a 15 percent stochastic demand variation, demonstrating acceptable robustness. The study identifies the absence of real hardware-in-the-loop evaluation and pedestrian modelling as limitations. Deployment recommendations target FCDA transport authority for a physical pilot at the Berger Roundabout as the highest-traffic test site in Abuja.

Keywords: reinforcement learning, adaptive traffic control, SUMO simulation, Abuja, deep Q-network