Joint Optimization of Vehicular Edge Computing and DeepJSCC Semantic Communication with Multi-Objective Awareness

Abstract

This article studies a Vehicular Edge Computing (VEC) system where connected vehicles generate image-based semantic tasks and process them locally or offload them to roadside units (RSUs). We jointly optimize RSU association, transmit power, local computing/edge offloading, and deep joint source-channel coding (DeepJSCC) tier selection, with semantic quality characterized by Cityscapes-based lookup tables between signal-to-noise ratio (SNR) and mean intersection over union (mIoU). A multi-objective reward prioritizes valid task completion, age of information (AoI), delay, energy efficiency, queue stability, and semantic accuracy. We implement an online joint model predictive control (MPC) policy that enumerates local and edge candidates in each slot and selects actions using a lexicographic objective. Experiments show that the proposed coordinated optimization achieves the best optimized reward, AoI, delay, and energy per successful task among B1-B5 and Ours.