Robust mean field game of autonomous vehicles on multilane roadways via attention-based adversarial networks

Abstract Drawing inspiration from the highly complex and nonlinear traffic flow phenomenon, first a model for the movement of autonomous vehicles (AVs) on multilane roadways, based on a robust mean field game (MFG) framework, is formulated. By considering the effect of adversarial disturbances in the traffic dynamics along with the effect of uncontrollable noise in the lateral movement of the vehicles, a min–max optimal control problem is converted to a coupled system of forward-in-time Fokker–Planck–Kolmogorov (FPK) and backwards-in-time Hamilton–Jacobi–Bellman–Isaacs (HJBI) equations. By formulating various cost functions, an equilibrium and a non-equilibrium (ML) model is developed, and their link with traffic flow is described. Second, a generative adversarial network (GAN)-based machine learning model is proposed as a surrogate tool for modelling the dynamics of the robust MFG framework. The GAN model is equipped with two physics-informed attention-based frameworks followed by a residual-based adaptive sampling strategy to enhance the learning efficiency and improve the accuracy. The validity of the model is tested against the synthetically generated traffic flow scenarios.