Interpretable Data-driven Driver Behaviour Modelling with Adversarial Inverse Reinforcement Learning

Understanding and predicting driving behaviours is essential for advancing traffic management, shaping transportation policy, and introducing new mobility technologies such as connected and autonomous vehicles. This study examines driving behaviour at two complementary levels: continuous operational car-following and network-wide discrete tactical route choice. We propose an interpretable data-driven modelling approach based on Adversarial Inverse Reinforcement Learning (AIRL) as a general framework that formulates a driver’s sequential decision making as a Markov Decision Process (MDP), applicable at both microscopic and macroscopic levels. AIRL jointly learns policies and reward functions from demonstration driving data, offering a principled way to uncover the motivations behind observed driving actions. However, existing applications rarely assess the consistency between learned policies and reward functions. To address this, we develop a structured five-step interpretability workflow that treats the recovered reward as an explicit utility model and validates its alignment with the induced policy. Empirical evaluations on real-world datasets show that the framework generates realistic and generalisable behaviours. In car-following, analyses highlight speed differences as dominant factors, while in route choice, road types and distance to the destination emerge as key determinants. Reward and policy signals are further compared, revealing cases of alignment and divergence that shed light on decision-making dynamics. Overall, the proposed AIRL-based framework achieves predictive accuracy on par with contemporary data-driven methods, while offering substantially greater interpretability through its explicit reward-policy decomposition. The extracted reward functions provide intuitive insights into latent behavioural motivations, and the application of interpretation methods clarifies the incentives shaping both short-term and long-term driving decisions. This work contributes a novel interpretable framework for driving behaviour modelling and offers a robust foundation for building safer, more transparent, and human-aligned mobility systems.