Maximum‐Entropy Reinforcement Learning for Online Computation Offloading in the IoT Integrated With Multi‐Access Edge Computing

Multi‐access edge computing (MEC) has emerged as a pivotal technology for enhancing the performance of edge networks. However, the continuous network tasks tend to lead to substantial data backlogs due to processing delays in dynamic MEC scenarios with stochastic data arrivals and time‐varying wireless channels. This can result in performance degradation or, in severe cases, network failure, causing great losses. In this paper, we propose a novel maximum‐entropy reinforcement learning (MERL)‐based online computation offloading (MEOCO) algorithm, which can maximize edge network processing capabilities and stabilize data queues without prior knowledge of random data arrivals and time‐varying wireless channels. The MEOCO algorithm leverages two deep neural networks as value functions to evaluate the policy network, significantly reducing the computational overhead and yielding valuable insights for future research. Moreover, MEOCO incorporates the entropy measure of the policy into the reward function to train a stochastic policy that can capture all potentially valuable decisions, thereby enabling it to make optimal offloading decisions. Extensive experiments demonstrate that our work performs comparably to the optimal algorithm and that it supports immediate deployment without repeated training, making it particularly suitable for implementation in dynamic MEC networks. The simulation code is accessible at https://github.com/Shanyang‐Cpu/MEOCO .