Task Offloading and Resource Allocation for ICVs in CFNs via Constrained MARL

Computing First Network (CFN) aims to enable unified scheduling of computing nodes across the network, positioning it as a promising paradigm for managing the computationally intensive tasks of Intelligent Connected Vehicles (ICVs). Given the highly dynamic vehicle network topology and uneven vehicle density, ICV tasks may fail due to limited computing resource on individual node. Thus, it is crucial for CFN to efficiently schedule I CV tasks and allocate resources to ensure their successful completion. However, existing task scheduling methods often struggle with large action spaces and tend to over-look the correlation between offloading and resource allocation decisions. Additionally, the high computing resource demands of ICV tasks require CFN to allocate resources within constrained range to ensure timely overall task completion. To address these challenges, we propose a Constrained and Action Represented Multi-Agent Reinforcement Learning (CARMARL) framework that generates reliable policies to meet the QoS requirements of I CV tasks while fully utilizing computing resources in CFN. Specifically, CARMARL maintains a Conditional Variational Autoencoder (CVAE) to reduce the action space and enable agents to learn the correlation between offloading and allocation decisions and a Reward Constrained Policy Optimization (RCPO) module to ensure resource allocation remains reasonableness. Extensive experiments demonstrate that CARMARL improves task completion rate by up to 25.2 % compared to all baselines.