Multiagent Deep-Reinforcement-Learning-Based Resource Allocation for Heterogeneous QoS Guarantees for Vehicular Networks

Vehicle-to-vehicle communications can offer direct information interaction, including security-centered information and entertainment information. However, the rapid proliferation of vehicles and the diversity of communications services demand for a more intelligent and efficient resource allocation framework to enhance network performance. In this article, a multi-agent deep reinforcement learning-based resource allocation framework is developed to jointly optimize the channel allocation and power control to satisfy the heterogeneous Quality-of-Service (QoS) requirements in heterogeneous vehicular networks. In the proposed framework, the utility maximization problem is formulated by considering two types of traffics, i.e., the strict ultrareliable and low-latency requirements for safety-centric applications and the high-capacity requirements for entertainment applications. The utility of each vehicular users is formulated as a multicriterion objective function by taking into account the heterogeneous traffic requirements. To overcome the drawbacks of the traditional totally centralized and distributed deep reinforcement learning-based resource allocation approaches, we propose a multi-agent deep deterministic policy gradient algorithm with centralized learning and decentralized execution to solve the formulated optimization problem. The normalization of the input states and reward functions is introduced to speed up the training and learning progress of the proposed algorithm. Simulation results show the superiority of the proposed algorithm in terms of the convergence and system performance through the comparison with the other methods and schemes for the delay-sensitive applications and delay-tolerant applications.