COORDINATED MULTIAGENT NEURAL NETWORK ARCHITECTURE FOR REAL-TIME DECISION MAKING IN AN INTELLIGENT NETWORKED ENVIRONMENT

Urban traffic congestion poses a persistent challenge in rapidly developing regions such as Port Har-court, Nigeria. Traditional methods often lack adaptability to dynamic urban road networks. In this paper, a coordinated multi-agent deep reinforcement learning (MADRL) system that leverages graph neural networks (GNNs) for optimizing traffic flow is presented. The methodology involves building a custom environment that mirrors the Road topology of Port Harcourt and deploying agents to minimize congestion across hotspots. These agents are trained using DQNs enhanced with graph layers to capture spatial dependencies between junctions. A dense variational Bayesian layer is used to improve Q-value approximation generalization. The custom environment replaces traditional OpenAI Gym simulations, allowing for direct dataset-based learning. The empirical results demonstrate the model’s robustness, achieving steady learning with low entropy (0.03) and high reward stability across episodes, resulting in steady reward growth