Vehicular Edge Intelligence: DRL-Based Resource Orchestration for Task Inference in Vehicle-RSU-Edge Collaborative Networks

Vehicular edge intelligence, distinct from traditional edge intelligence, exhibits unique characteristics, including the mobility of vehicles, uneven spatial and temporal distribution of vehicles, and variability in the AI models deployed on vehicles, Roadside Units (RSUs), and edge servers (ESs). In this paper, we propose a Deep Reinforcement Learning (DRL)-based resource orchestration scheme for task inference in vehicle-RSU-edge collaborative networks. In our approach, vehicles' inference tasks can be processed on the vehicles, RSUs, or ESs, encompassing a total of 9 possible scenarios based on the cross-RSU mobility of vehicles. The scheme jointly optimizes task processing decision-making, transmission power allocation, computational resource allocation, and transmission rate allocation. The objective is to minimize the total cost, which involves a trade-off between task processing latency, energy consumption and inference error rate across all vehicle tasks. We design a DRL algorithm that decomposes the original optimization problem into sub-problems and efficiently solves them by combining the Softmax Deep Double Deterministic Policy Gradients (SD3) algorithm with multiple numerical methods. We analyzed the complexity and convergence of the algorithm. Specifically, we demonstrated its low complexity and fast, stable convergence, which prove its effectiveness in solving the problem. And we demonstrate the superiority of our scheme by comparing it with 5 benchmark schemes across 6 different scenarios.