Federated Deep Reinforcement Learning Framework for Intelligent Traffic Management Using AI-Driven IoT Edge Systems

In the modern urban context, traffic congestion is a significant problem that can lead to a range of issues, including longer commutes, higher fuel consumption, pollution, and economic losses. This research proposes a Federated Deep Reinforcement Learning (FDRL) approach for intelligent traffic control based on AI enabled edge systems in IoT. The proposed system is based on Internet of Things (IoT) devices, Edge Computing, Deep Q-Network (DQN) algorithm and Federated Learning, which will allow for decentralized, privacy-preserving, and adaptive traffic signal optimization. The Nigeria Road Traffic Data dataset was pre-processed by cleaning the data, engineering features and normalizing the data with Z score before distributing the data to various edge nodes for local model training. This reduces the need for the DQN agents to learn optimal traffic signal control policies and the Federated Averaging (FedAvg) algorithm was used to merge local model parameters into a global model without sharing traffic information. The proposed framework was tested and analysed based on metrics like cumulative reward, training loss, average vehicle waiting time, queue length, traffic throughput and communication overhead. The results revealed that the training reward of DQN has been improved from 52.4 to 294.5, and the training loss has been reduced from 1.25 to 0.11, which means the model has converged successfully. The average vehicle waiting time decreased by 57.96%, the length of the queues decreased by 61.01%, and traffic throughput increased by 49.81% on average. In addition, this resulted in about 76.8% reduction in the overhead of communication as compared to the centralized learning approach. The results prove that the proposed FDRL framework can achieve optimal traffic flow and alleviate traffic congestion, while ensuring data privacy and making the ITS more scalable.