SAC-Based Collaborative Task Offloading and Path Optimization for Vehicular Edge Computing

Abstract

With the rapid development of Vehicle-to-Everything (V2X) and Vehicular Edge Computing (VEC), the massive computation-intensive tasks generated by intelligent vehicles during driving, including environmental perception, path planning, and autonomous driving decision-making, impose extremely high requirements on real-time performance and reliability. Traditional task offloading strategies often consider communication quality or computing resources in isolation, thereby neglecting the inherent coupling between a vehicle’s dynamic driving path and its offloading decisions. This leads to low task completion rates, long travel times, and insufficient utilization of edge computing resources. Therefore, addressing the joint optimization problem in VEC scenarios, this paper proposes a deep reinforcement learning (DRL) algorithm based on an improved hybrid action space Soft Actor-Critic (SAC-Discrete). An agent capable of simultaneously handling continuous speed control and discrete offloading decisions is designed, enabling it to learn to collaboratively trade off task offloading and path optimization in dynamic traffic environments. This paper systematically compares the proposed algorithm with five baseline methods: greedy strategy, fixed path optimization, fixed offloading optimization, random strategy, and a discretized-action version of Vanilla SAC. Experimental results demonstrate that our algorithm converges rapidly and achieves high stability, and the proposed joint optimization framework significantly improves system efficiency and resource utilization while reducing average travel time.