Abstract
The increase in Internet of Things (IoT) devices has increased demand for effective and reliable task offloading strategies in edge-computing environments. These systems struggle to balance limited computational resources, varying network conditions, and complex dependencies between subtasks. To address these challenges, this article developed a task offloading framework using the Deep Deterministic Policy Gradient (DDPG) algorithm with Uncertainty-aware Prioritized Experience Replay (UPER) and Dual Critic Attention (DCA) mechanism. The task offloading issue is modeled as a Markov Decision Process (MDP), in which agents learn optimal policies to minimize delay and energy consumption while ensuring high task success rates. The UPER module enhances learning efficacy by sampling transitions with high epistemic uncertainty and temporal difference errors, allowing the DCA component to dynamically balance energy and latency objectives using dual-critic networks. The simulation results show that the proposed model outperforms existing algorithms in terms of offload success rate and minimization of average delay and energy consumption.