Optimizing Traffic Light Control in Developing Cities Using Adaptive Federated Reinforcement Learning in Vehicular Ad Hoc Networks

Efficient control of urban intersections remains a major challenge in developing countries, where detection infrastructures are scarce or unreliable. This paper introduces a novel distributed architecture for traffic light management that relies exclusively on vehicle to infrastructure and infrastructure to infrastructure communications, avoiding the use of physical sensors embedded in the roadway or cameras. The approach integrates three complementary components: a long short-term memory model for traffic flow prediction, a deep q-learning agent for local phase optimization, and an adaptive federated learning mechanism ensuring coherence and robustness among traffic lights. The proposed framework was validated through simulations on the real-world “Ancien Dalip” intersection in Douala, Cameroon, using SUMO coupled with OMNeT++ and Veins. Results demonstrate significant improvements over baseline methods (fixed-time and deep q-learning), including reduced congestion, lower CO \(_2\) , NO \(_x\) , and CO emissions, enhanced fairness across competing flows, and over 50% reduction in emergency vehicle waiting times. These findings highlight the relevance of combining predictive, reinforcement, and federated learning to design sustainable and scalable traffic management solutions tailored to resource-constrained urban environments.