Predictive Latency-Aware Federated Deep Reinforcement Learning for Adaptive Task Scheduling in IoT-Enabled Fog Computing

Abstract

The increasing deployment of latency-sensitive Internet of Things (IoT) applications has intensified the need for intelligent task scheduling mechanisms in fog computing environments. Conventional scheduling approaches, including heuristic and centralized machine learning techniques, often fail to adapt to dynamic workload variations and mobility-induced network changes, resulting in increased task latency. This paper proposes a Predictive Latency-Aware Federated Deep Reinforcement Learning (PLA-FDRL) framework for adaptive task scheduling in IoT-enabled fog networks. The proposed framework integrates latency prediction, mobility-aware fog node selection, and federated deep reinforcement learning to proactively allocate tasks to optimal fog resources. Each fog node independently trains a Deep Q-Network (DQN) using local observations and periodically participates in federated aggregation without sharing raw data. A latency-aware reward function jointly minimizes transmission, queueing, processing, and migration delays. Experimental evaluation under dynamic IoT workloads demonstrates significant reductions in average task latency and response time compared with FCFS, Round Robin, centralized DQN, and conventional federated reinforcement learning schedulers. Results indicate that the proposed framework improves responsiveness and scalability while preserving data privacy.