Reinforcement Learning for Accident Risk-Adaptive V2X Networking

The significance of vehicle-to-everything (V2X) communications has been ever increased as connected and autonomous vehicles (CAVs) get more emergent in practice. The key challenge is the dynamicity: each vehicle needs to recognize the frequent changes of the surroundings and apply them to its networking behavior. This is the point where the need for machine learning is raised. However, the learning itself is extremely complicated due to the dynamicity as well, which necessitates that the learning framework itself must be resilient and flexible according to the environment. As such, this paper proposes a V2X networking framework integrating reinforcement learning (RL) into scheduling of multiple access. Specifically, the learning mechanism is formulated as a multi-armed bandit (MAB) problem, which enables a vehicle, without any assistance from external infrastructure, to (i) learn the environment, (ii) quantify the accident risk, and (iii) adapt its backoff counter according to the risk. The results of this paper show that the proposed learning protocol is able to (i) evaluate an accident risk close to optimal and, as a result, (ii) yield a higher chance of transmission for a dangerous vehicle.