Machine Learning for Optimizing Autonomous Vehicle Communication Protocols

Abstract

Effective dissemination of accurate data is crucial to enhance driving efficiency and safety in intelligent transportation systems. Connected vehicles communicate with the transportation network and the vicinity of nearby vehicles to enhance road safety and improve mobility. Current advancements in the automotive industry have seen the development of highly advanced systems that rely on wireless communication, predictive modeling, and 3D mapping. The key to successful cooperation of vehicle-to-everything, vehicle-to-vehicle, vehicle-to-pedestrian, and vehicle-to-infrastructure communication lies in part on the radio protocols used and their efficiency. Cooperative intelligent transportation protocols such as Dedicated Short Range Communication and cellular long-term evolution are used to relay packets, but often result in unwanted collisions and congestion with the potential to be detrimental.

One mode of improvement of these standard radio protocols is formulated through machine learning. This paper focuses on performing reinforcement learning approaches, deep Q-learning, and Q-learning to optimize an operation regarding service data unit distances, maximum contention window, and retransmission limit. The paper also suggests a recurrent neural network that can be used as an enhancement to the already proposed deep Q-learning framework presented throughout the research. Additionally, we suggest that parameter metrics can also be varied, such as vehicle-to-vehicle distance, traffic conditions, or antenna gain, to analyze the relative efficiency of these new machine learning approaches over the classic model.