Behavior Cloning With Fuzzy Logic for Accelerating Early-Stage DRL in Edge Offloading

Deep Reinforcement Learning (DRL) has proven effective in addressing the complexities of task offloading in heterogeneous and dynamic edge computing environments. However, the iterative training process of DRL, particularly in its early stages, is computationally intensive and time-consuming, posing significant challenges for delay-sensitive applications. Furthermore, DRL agents are typically initialized with random weights, as is common in deep learning, resulting in models that lack meaningful behavioral priors and further complicate the learning process. To tackle these issues, this paper leverages fuzzy logic to improve the early-stage performance of DRL agents in task offloading through behavior cloning. By exploiting the high-level reasoning capabilities of fuzzy logic, behavior cloning equips DRL agents with heuristic knowledge, thereby reducing exploration costs and accelerating training in the early learning stages. To facilitate the seamless integration of fuzzy logic–based knowledge into DRL agents, five distinct behavior cloning strategies are introduced: only cloning, norm cloning, fading cloning, runtime cloning, and hybrid cloning. Extensive simulations in multi-access edge computing environments with multiple edge servers and user devices demonstrate that integrating fuzzy logic with DDQN-based reinforcement learning via behavior cloning significantly improves the Average Task Completion Rate (ATCR) and reduces the Average Task Processing Time cost (ATPT) of all user devices over learning episodes during early training stages, while concurrently enhancing the stability of learning progress compared to DRL baseline approach.