Transformer-guided optimization of railway alignments using an adaptive rapidly-exploring random tree algorithm

Abstract

Railway alignment design is a crucial part of a railway project. Despite the widespread success of computer-aided alignment optimization methods in determining alignments, effectively exploring the objective function’s descent direction (OFDD) remains challenging, particularly when navigating complex alignment search spaces. To address this issue, it is essential to comprehensively consider key factors, including the global and local environment, explored and unexplored search spaces, as well as established alignment search strategies and potential new ones that may emerge during the OFDD optimizing process. Therefore, an alignment-oriented Transformer framework is formulated in this work. In this framework, various real-world railway cases are input into a stacked Transformer framework to learn an optimized OFDD strategy. Specifically, the model handles regular inputs (i.e., global and local contexts, long-term goals) and irregular inputs (i.e., historical paths) using two separate stacked Transformer encoders. Afterward, an adaptive rapidly-exploring random tree star (Ada-RRT-star) method is developed by integrating the Transformer framework’s output to guide RRT’s search direction as well as to enhance the solution quality. Ultimately, the proposed method is applied to a realistic railway case, where the results demonstrate its superiority over the conventional 3D-RRT-star algorithm in terms of solution quality. Besides, the best alignment generated by the Ada-RRT-star also outperforms the manually-designed alignment.