Deep Reinforcement Learning-Based Cloud-Edge Offloading for WBANs

Wireless body area networks (WBANs) are essential for real-time health monitoring using physiological sensors, but these sensors have limitations in computing power, storage, and battery life. To address these issues, we propose a cloud-edge computing architecture where data from sensors are initially processed by WBAN edge devices (smartphones, PDAs) before being offloaded to cloud servers for extensive analysis. We model the offloading decision as a Markov decision process (MDP) and design a system utility function that balances processing latency, power consumption, and the quality of communication. Traditional deep reinforcement learning algorithms like deep deterministic policy gradient (DDPG) optimize MDPs but face slow convergence and local optima issues. To overcome these, we introduce a multi-step deep deterministic policy gradient (MSDDPG) algorithm that uses a multi-step update mechanism, enhancing learning efficiency and solution accuracy. Experimental results show that MSDDPG significantly reduces processing delay and energy consumption while improving system utility. Additionally, our approach considers WBAN edge device constraints, ensuring practical applicability, extending device battery life, and enhancing performance in real-world scenarios.