Objective-directed deep graph generative model for automatic and intelligent highway interchange design

Abstract

Highway interchanges have traditionally been designed through a time-consuming manual process. To enhance efficiency and effectiveness, this paper develops an objective-directed automatic and intelligent interchange design method using graph conditional variational autoencoder. Based on interchange graph representation and augmentation techniques, data are collected from diverse interchanges types and converted into graphs that store design parameters. Aiming at graph reconstruction and fitting data distribution, proposed model learns to generate optimized interchanges by embedding design objectives including throughput and total ramp length. For evaluation, predictors are used to directly output interchange properties, enabling the quick screening of structures. Results demonstrate significant improvements with generated designs showing up to 7.67 % increased throughput and 27.63 % reduced total ramp length compared to traditional methods. The generated set contains a high proportion of valid, novel and unique interchanges. These advancements highlight the potential for generative model in creating more efficient and valid interchanges.