Joint RAN Slicing and Computation Offloading for Autonomous Vehicular Networks: A Learning-Assisted Hierarchical Approach

In this paper, a two-timescale radio access network (RAN) slicing and computing task offloading problem is investigated for a cloud-enabled autonomous vehicular network (C-AVN). We aim at jointly maximizing the communication and computing resource utilization with diverse quality-of-service (QoS) guarantee for autonomous driving tasks. Specifically, to capture the small-timescale network dynamics, a computing task scheduling problem is formulated as a stochastic optimization program, for maximizing the long-term network-wide computation load balancing with minimum task offloading variations. Due to the large problem size and unavailable network state transition probabilities, we employ cooperative multi-agent deep Q-learning (MA-DQL) with fingerprint to solve the problem by learning the set of stationary task offloading policies with stabilized convergence. Given the task offloading decisions, we further study a RAN slicing problem in a large timescale, which is formulated as a convex optimization program. We focus on optimizing the radio resource slicing ratios among base stations, to maximize the aggregate network utility with statistical QoS provisioning for autonomous driving tasks. Based on the impact of radio resource slicing on computation load balancing, we propose a two-timescale hierarchical optimization framework to maximize both communication and computing resource utilization. Extensive simulation results are provided to demonstrate the effectiveness of the proposed framework in comparison with state-of-the-art schemes.

Automated summary

This paper investigates a two-timescale radio access network (RAN) slicing and computing task offloading problem for a cloud-enabled autonomous vehicular network (C-AVN). The problem is addressed by formulating a stochastic optimization program for small timescale network dynamics, and employing cooperative multi-agent deep Q-learning (MA-DQL) for stabilization convergence. Through extensive simulation results, the proposed framework demonstrates effectiveness in maximizing communication and computing resource utilization.