Risk-Constrained Learning of Personalized Driving Behaviors for Autonomous Vehicles

Achieving intelligent driving assistance hinges on a precise understanding and prediction of human driving behavior, which requires reliable and interpretable behavior models. However, existing neural network–based approaches have limited capability in capturing drivers’ latent intentions and personalized preferences, and their lack of interpretability constrains prediction accuracy and the adaptability of assistive strategies. Human driving behavior can be formalized as a Markov Decision Process (MDP), integrating goal pursuit, cognitive risk perception, and interaction intentions with surrounding vehicles. To address these challenges, this study proposes a personalized driving behavior modeling framework based on inverse reinforcement learning (IRL), enabling the inference of individual reward functions and the characterization of personalized behavioral traits. The framework first establishes a driver cognitive model grounded in the concept of a risk field, describing the driver’s subjective assessment and interaction tendencies toward surrounding vehicles in terms of risk perception and courtesy intention. This formulation quantifies individual risk preferences and safety boundaries. A personalized reward structure is then constructed by incorporating behavioral features, and the driver’s implicit preferences are recovered through IRL. By extracting state–action features from naturalistic driving data, the framework infers reward parameters that represent intrinsic behavioral motivations, enabling interpretable and data-driven modeling of personalized driving behavior. Finally, the learned personalized reward model is integrated into a reinforcement learning framework as a guidance signal, allowing the agent to generate adaptive, human-like driving behaviors consistent with individual psychological characteristics and preferences. Experimental results on naturalistic highway data verify that the proposed approach effectively reproduces diverse driving styles and enhances behavioral interpretability.