Cooperative Spectrum Sensing Approach in C-V2X based on Multi-Agent Reinforcement Learning

Abstract

In cellular vehicle-to-everything (C-V2X) Mode 4, the autonomous mode is based on spectrum sensing for vehicles selecting spectrum resources, however, the spectrum sensing process may be imprecise, leading to frequent packet collisions especially when vehicles are in the congested scenario with high speeds. In this paper, we propose a cooperative spectrum sensing approach based on multi-agent reinforcement learning (MARL) to improve the spectrum sensing accuracy. The proposed algorithm includes the sense channels selection procedure and the cooperative users selection procedure. Specifically, we imitate the process of vehicles selecting sense channels as Indian buffet process (IBP) to predict the channel selection probability, and establish the belief scheme of cooperative users based on the historical sensing results, then exploit deep Q-learning network (DQN) to select sense channels based on the channel selection probability and select cooperative users based on the belief. Simulation results show that the proposed algorithm can improve success sensing rate and reducing collision rate compared with other cooperative and non-cooperative methods.