Modelling and interpreting evacuation time and exit choice for large-scale ancient architectural complex using machine learning

The intricate spatial structures of large-scale ancient architectural complexes, combined with their complex crowd compositions, hinder evacuation efficiency. However, research on the evacuation of such complexes is limited. Moreover, among the existing evacuation studies, most solely rely on either simulation models or machine learning techniques. Few integrate the strengths of simulation modeling in capturing otherwise inaccessible evacuation data with the advantages of machine learning in addressing complex problems. This study aims to utilize random forest algorithm, based on Pathfinder simulation evacuation data, to model the evacuation time and exit choice of occupants within such complexes. Additionally, we employ tools to interpret the random forest models, analyzing the influence of various factors on evacuation time and exit choice. Specifically, this methodology was applied to examine evacuation time and exit choice of 4800 occupants, comprising both tourists and residents, within the Xidi ancient architecture complex that spans nearly 130,000 square meters. The reliability of the random forest evacuation models is validated using real experimental data. The results indicate that the impact of environmental factors on evacuation time is often nonlinear. Compared to environmental factors, individual factors tend to have a lesser influence on evacuation time. For exit choice, the influence of environmental factors varies in significance at different scales. Individual factors have a notably higher influence on exit choice than on evacuation time. We also discern that evacuation distance affects the degree to which other factors influence exit choice. Overall, this research contributes to a deeper understanding of evacuation time and exit choice in such complexes, offering decision-makers more comprehensive evacuation strategies.