Task Offloading via Prioritized Experience-Based Double Dueling DQN in Edge-Assisted IIoT

In the Industrial Internet of Things (IIoT), Multi-access Edge Computing (MEC) emerges as a transformative paradigm for managing computation-intensive tasks, where task offloading plays an important role. However, due to the complex environment of IIoT, existing deep reinforcement learning-based schemes suffer from significant shortcomings in accuracy and convergence speed during model training when addressing the issue of task offloading. In this paper, to solve this problem, we propose an online task offloading scheme based on reinforcement learning, leveraging the double deep Q network (DQN) and dueling DQN with a prioritized experience replay mechanism, called the <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">P</b>rioritized experience-based <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">D</b>ouble <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">D</b>ueling <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">DQN</b> task offloading scheme (P-D3QN). P-D3QN enhances action selection accuracy using double DQN and mitigates Q-value overestimation by decomposing state and advantage using dueling DQN. Additionally, we adopt the prioritized experience replay mechanism to enhance the convergence speed of model training by selecting transitions that induce a higher training error between the evaluation network and the target network. Experimental results demonstrate that P-D3QN outperforms several state-of-the-art schemes, achieving a reduction of 21.0% in the average cost of the task and improving the completion rate of the task by 19.5%.