Intelligent Task Offloading in IoT-Driven Digital Twin Systems via Hybrid Federated and Reinforcement Learning

Efficient resource allocation in IoT-driven Digital Twin (DT) systems was crucial for ensuring reliable and timely task processing in dynamic environments. In this study, we proposed an advanced task offloading strategy, FLaMAD (Federated Learning and Multi-Agent Deep Reinforcement Learning), to optimize performance metrics across various datasets. FLaMAD leveraged hybrid Federated Learning (FL) for decentralized model training, enhancing data privacy, and Multi-Agent Deep Reinforcement Learning (MADRL) for adaptive task offloading decisions. The approach integrated seamlessly with IoT-LAB, OpenEdge, and TAPAS Cologne datasets, providing insights into device data, edge resource profiles, and mobility patterns within smart city infrastructures, including vehicular networks (IoV) and roadside units (RSUs). Simulation results demonstrated substantial improvements over baseline methods: FLaMAD achieved a task completion rate (TCR) of 95% on IoT-LAB, 94.5% on OpenEdge, and 96.1% on TAPAS Cologne. Compared to traditional approaches, FLaMAD reduced energy consumption by approximately 15% to 18% (350 J to 360 J), decreased latency by 25% (average of 120 ms), and optimized resource utilization with edge and cloud servers operating at 85% efficiency.