KGDRL-HAO: A knowledge reasoning-guided deep reinforcement learning framework for highway alignment optimization

Journal article, 2026

Highway geometric design is a continuous alignment optimization process involving geometric constraints and multi-objective trade-offs. However, current highway alignment optimization methods are often limited by fragmented and unstructured knowledge representations, and the lack of prior knowledge guidance results in poor exploration efficiency and restricted practical applicability . To address these issues, this study proposes a knowledge reasoning-guided deep reinforcement learning framework for highway alignment optimization, named KGDRL-HAO. The KGDRL-HAO employs graph neural networks based on multi-hop inference and large language models with contextual understanding to capture semantic dependencies among design variables, constraint conditions, and optimization objectives, outperforming baseline knowledge reasoning models. Then, by incorporating structured prior knowledge into the soft actor-critic learning process, the KGDRL-HAO enhances exploration efficiency and decision consistency, reducing the weighted total cost by 6.4% while satisfying design constraints. The findings would facilitate intelligent and interpretable highway geometric design through knowledge-augmented optimization and natural language interaction.