Multiple Channel Access using Deep Reinforcement Learning for Congested Vehicular Networks

Vehicular Ad-hoc Network (VANET) is a standard protocol for wireless vehicular communication that enables Vehicle to Vehicle (V2V) and Vehicle to Infrastructure (V2I) communications. VANET safety applications aim to prevent traffic accidents and require a high Packet Delivery Ratio (PDR) and low latency of safety packet broadcast. When a large number of vehicles simultaneously access a limited channel resource for the safety broadcast, the safety requirements impose more challenges; the communication performance will significantly degrade due to network congestion. Especially, infrastructureless VANETs, which only allow V2V communication, vehicles are supposed to overcome the congestion problem using a self-adaptation scheme without the aid of infrastructures. In this paper, we propose a self-adaptive MAC layer algorithm employing Deep Q Network (DQN) with a novel contention information-based state representation to improve the performance of the V2V safety packet broadcast. The proposed algorithm operates a fully distributed manner, and it is evaluated by simulations considering various levels of traffic congestion.