Graph attention network for Car-Following Model under game between desired and real state

Abstract

The Car-following model plays a major role in the study of traffic flow theories and traffic simulation. Car-following behaviours can be regarded as a game process between the desired driving state and the real traffic state. By defining the expected car-following state, this paper expresses the game process with a graph attention neural network, and developed the GATCF car-following model. Different from other models that apply time series information as input, GATCF only needs instantaneous information as feature input. Two simulation cases, trajectory sequence prediction and multi-vehicle simulation, are performed with the I-80 data in the NGSIM project. Compared with other car-following models such as Gipps model, Optimal Velocity model, and Intelligent Driver model, the GATCF model developed in this paper shows higher accuracy and stability.