Deep Deterministic Policy Gradient to Minimize the Age of Information in Cellular V2X Communications
Zoubeir Mlika, Soumaya Cherkaoui
Abstract
This paper studies the problem of minimizing the age of information (AoI) in cellular vehicle-to-everything communications. To provide minimal AoI and high reliability for vehicles’ safety information, non-orthogonal multiple access is exploited. We reformulate a resource allocation problem that involves half-duplex transceiver selection, broadcast coverage optimization, power allocation, and resource block (RB) scheduling. First, to obtain the optimal solution, we formulate the problem as a mixed-integer nonlinear programming problem and then study its NP-hardness. The negative result of NP-hardness motivates us to design efficient sub-optimal solutions. Consequently, we model the problem as a single-agent Markov decision process (MDP). The MDP model helps in solving the problem efficiently using fingerprint deep reinforcement learning (DRL) techniques such as deep-Q-network (DQN) methods. Nevertheless, applying DQN is not straightforward due to the curse of dimensionality implied by the large and mixed action space that contains discrete RB scheduling decisions and continuous power and coverage optimization decisions. Therefore, to solve this mixed discrete/continuous problem efficiently simply and elegantly, we propose a decomposition technique that consists of first solving the discrete subproblem using a matching algorithm based on state-of-the-art stable roommate matching and then solving the continuous subproblem using DRL algorithm that is based on deep deterministic policy gradient (DDPG). We validate our proposed method through Monte Carlo simulations where we show that the decomposed matching and DRL algorithm successfully minimizes the AoI and achieves almost 66% performance gain compared to the best benchmarks for various vehicles’ speeds, transmission power, or packet sizes. Further, we prove the existence of an optimal value of broadcast coverage at which the learning algorithm provides the optimal AoI.