DRL-based dependent task offloading with delay-energy tradeoff in medical image edge computing

Task offloading solves the problem that the computing resources of terminal devices in hospitals are limited by offloading massive radiomics-based medical image diagnosis model (RIDM) tasks to edge servers (ESs). However, sequential offloading decision-making is NP-hard. Representing the dependencies of tasks and developing collaborative computing between ESs have become challenges. In addition, model-free deep reinforcement learning (DRL) has poor sample efficiency and brittleness to hyperparameters. To address these challenges, we propose a distributed collaborative dependent task offloading strategy based on DRL (DCDO-DRL). The objective is to maximize the utility of RIDM tasks, which is a weighted sum of the delay and energy consumption generated by execution. The dependencies of the RIDM task are modeled as a directed acyclic graph (DAG). The sequence prediction of the S2S neural network is adopted to represent the offloading decision process within the DAG. Next, a distributed collaborative processing algorithm is designed on the edge layer to further improve run efficiency. Finally, the DCDO-DRL strategy follows the discrete soft actor-critic method to improve the robustness of the S2S neural network. The numerical results prove the convergence and statistical superiority of the DCDO-DRL strategy. Compared with other algorithms, the DCDO-DRL strategy improves the execution utility of the RIDM task by at least 23.07, 12.77, and 8.51