A Stackelberg-guided deep Q-network framework for risk-aware decision-making in autonomous highway driving

Real-time decision-making in high-speed highway driving requires autonomous agents to balance operational efficiency and safety amid uncertain multi-vehicle interactions. This paper presents a lightweight, modular decision-making framework integrating Stackelberg-based risk filtering with a hierarchical Deep Q-Network (DQN). The Stackelberg module estimates potential follower reactions and pre-filters unsafe actions, while the DQN selects maneuvers from the remaining safe set. A model predictive control (MPC) module, included for trajectory feasibility, ensures low-level physical compliance. Experiments conducted in CARLA ’ s Town04 environment demonstrate that the proposed architecture reduces collision rates and produces smoother driving behavior in complex traffic flows, without compromising average speed. These findings support the viability of combining game-theoretic foresight with reinforcement learning in real-time highway scenarios.