An Interpretable Deep Learning Decision-Making Framework for Autonomous Lane-Change Inspired by Human Driving Experiences

Lane change plays a critical role in autonomous driving and directly affects traffic safety and efficiency. Although deep learning-based lane-change decision-making frameworks have achieved promising results, they still face fundamental challenges in producing human-consistent and trustworthy behavior, mainly due to: 1) Inadequate psychology-informed personalization, as most frameworks focus on physical variables but neglect psychological factors (e.g., risk tolerance, urgency), limiting their ability to capture individual differences in lane-change motivations. 2) Limited holistic understanding of traffic context, most frameworks lack consideration of high-level and interpretable indicators (e.g., traffic pressure) in comprehensively assessing dynamic traffic scenarios, limiting their capacity for human-like contextual understanding. 3) Lack of transparent and interpretable decision logic, as many frameworks operate as black boxes with opaque reasoning processes, hindering human-aligned explanation, weakening user trust, reducing accident traceability, and impeding model refinement. To this end, a policy-oriented contextual-reasoning fuzzy neural network (POCR-FNN) is proposed as a deep learning-based decision-making framework for personalized and interpretable autonomous lane-change. First, we develop a psychology-informed driving style classification by learning distinct fuzzy membership functions to enable style-specific policy learning. Second, we design a human-inspired local interaction-aware module that estimates traffic tension by combining interaction salience and contextual risk, enhancing contextual understanding. Finally, we integrate fuzzy logic with a deep learning-based policy network to enable rule-level decision reasoning with real-time interpretability and transparent traceability. Extensive experiments on multiple public highway and urban datasets demonstrate that POCR-FNN achieves state-of-the-art performance while significantly improving personalization and interpretability across various driving styles and scenarios.