Adaptive Task Scheduling in Fog Computing Using Federated DQN and K-Means Clustering

Task scheduling in distributed cloud and fog computing applications must be efficient to optimize resource utilization, minimize latency, and comply with strict service level agreements. The dynamic and heterogeneous of fog computing challenges such as balancing performance with variable task loads, resource constraints, and energy efficiency. This paper introduces a new Federated Learning Deep Queue-Learning (FLDQN) framework that integrates reinforcement learning with federated learning to overcome these challenges. The FLDQN approach enables decentralized task scheduling because fog nodes can train their local DQN models on localized data, merging their experiences to form a global model without sensitive data sharing. It further improves task prioritization by classifying tasks based on their execution time and deadlines, ensuring that high-priority tasks are scheduled first, which reduces SLA violations and task rejection rates. The proposed model dynamically adapts to the changing conditions of the fog environment through iterative learning that continuously improves work allocation. Substantial experiments conducted on large and small sizes of dataset show that the proposed FLDQN outperforms others, including standalone DQN and graph-based GGCN. For all dataset sizes small, medium, and large datasets, it reduces makespan by up to 30%, improves throughput, and reduces energy by distributing tasks on the most efficient node based on current system states. The results demonstrate the ability of FLDQN to optimize task execution in real time, while addressing the scalability and privacy issues inherent in fog computing. Combining federated and reinforcement learning proposed framework provides a flexible solution for distributed task scheduling, which is well suited for latency-sensitive applications in industrial automation, healthcare monitoring, and smart city deployments.