A Game Theoretic Four-Stage Model Predictive Controller for Highway Driving

Abstract

We develop a game theoretic model predictive controller (GTMPC) for autonomous driving in highway traffic. The hierarchical GTMPC uses game theory as the basis for its high-level controller by continuously playing games with the surrounding vehicles, which we call game candidate vehicles (GCV), to evaluate options. We pose the lane change situation in highway driving as Stackelberg game, where subject vehicle's (SV) strategies are the proper times to initiate/complete a lane change and the corresponding trajectories, and GCV's strategies are defined as accelerations and lane decisions. To capture SV's actual on-road benefit, we define SV's payoff as the negative of the cost function of the model predictive controller (MPC). GCV's payoff considers three factors including her position, headway and speed. A four-stage hybrid MPC is established as the low-level controller that controls both SV's longitudinal position and lane decision. To validate the effect of the controller, we implemented it into a virtual highway environment built in MATLAB SIMULINK. We first tested the controller's performance in a normal driving scenario against programmed traffic vehicles. Then we conducted a human-in-the-loop simulation in a mandatory lane change (MLC) scenario. The simulations showed that GTMPC is able to well predict surrounding vehicle's behavior during the interaction and make reasonable decisions in different situations even at the presence of human driver.