Pruning-based Deep Reinforcement Learning for Task Offloading in End-Edge-Cloud Collaborative Mobile Edge Computing

Abstract

The task offloading of Mobile Edge Computing (MEC) brings infinite possibilities to compute-intensive and latency-sensitive mobile applications. However, the dynamic nature of MEC systems and complex dependencies between computational tasks pose significant challenges to offloading decisions. In this paper, we address the task offloading problem of end-edge-cloud collaborative computing in MEC with task dependencies. Initially, we model inter-task dependencies using Directed Acyclic Graphs (DAG) and propose a task priority queue model to transform the DAG task model into a sequential queue model for easier task scheduling. Subsequently, we formulate a resource-constrained minimization problem for execution delay and energy consumption optimization. To tackle this problem, we introduce a Pruning-based Deep Reinforcement Learning algorithm (PR-DRL) to learn the intricate dependencies between MEC systems and subtasks for optimal offloading decisions. Specifically, PR-DRL incorporates a pruning function that enables the agent to focus on high-probability actions during training while filtering out low-probability actions, thereby achieving rapid algorithm convergence. Simulation results demonstrate that the proposed PR-DRL method outperforms traditional Deep Q Network methods both in terms of convergence speed and offloading performance, surpassing six other baseline task offloading methods.