Enhanced Task Offloading in Mobile Edge Computing: A Hybrid Approach Using Deep Q-Learning from Demonstrations and Heuristic Optimization

The increasing demands of delay-sensitive and resource-intensive applications in the IoT era have made Mobile Edge Computing (MEC) a promising solution for task offloading to nearby edge servers. Effective offloading in MEC requires strategies that balance energy consumption and meet strict deadline constraints to support the limited battery life of IoT devices and ensure low latency. Traditional heuristic algorithms, like DECO, efficiently manage energy and deadlines through predefined rules, while reinforcement learning (RL) methods adapt dynamically in changing environments. This paper proposes a hybrid model, DQfD-DECO, which combines the advantages of DECO and Deep Q-learning from Demonstrations (DQfD). By leveraging demonstration-based learning, our approach enhances initial offloading decisions and refines them through adaptive reinforcement learning. Comparative performance analysis in dynamic MEC scenarios reveals that DQfD-DECO achieves a superior balance of energy efficiency and deadline satisfaction, outperforming traditional rule-based methods in adaptability to fluctuating conditions.