Attentional value-factorization-based resource allocation and performance evaluation for intelligent connected vehicles

Intelligent connected vehicles are integrated equipment with communication, computing, and control capabilities. The diversity of communication demands and high-speed mobility of intelligent connected vehicles leads to the challenge of wireless resource allocation and performance evaluation. To manage wireless resources efficiently in complex environments, this paper proposes an attentional value factorization (AVF) based cooperative resource allocation and performance evaluation method, which is built on top of the actor-critic-based multiagent deep reinforcement learning (MADRL) framework. Specifically, AVF is constructed with hierarchical and heterogeneous critics to accurately evaluate the performance and then fed back to the resource allocation policy. The individual task-specific critics and the global critic are exploited to trade off local optimum against global optimum. In addition, a meticulously designed attentional action-value mixing network is used by the global critic to assign different credits for individual critics, factorizing the overall reward to each agent and further to each agent’s sub-tasks. Extensive experimental results show that AVF can promote cooperative resource allocation among agents by evaluating which agents and what actions have been contributing most to the overall quality of communication, achieving optimal resource efficiency and accurate evaluation.