Calibrating parameters of crowd evacuation simulation at strategic, tactical and operational levels: Which one matters most?

Abstract

Background: Simulating a process of crowd evacuation using an agent-based model requires modellers to specify values of a whole range of parameters each determining certain aspects of evacuee behaviour. While potential sensitivity of simulation outputs to each single parameter is undoubted, one critical question is that to which class of parameters such numerical models are most sensitive when estimating evacuation times? The question is of significance given that available data and resources for calibration are limited and priorities often need to be made. Methods: Here, we employ a simulation model of evacuation comprised of multiple integrated layers. After providing an overview of the calibration procedure for key parameters, we perform extensive sensitivity analyses in complex simulated spaces with large crowds to answer the above question. Findings: Our analyses showed that, once parameters are contrasted on a consistent and comparable scale, estimates of simulated evacuation time are, by far margins, most sensitive to the value of mechanical movement parameters at locomotion layer, those that determine the discharge rate at bottlenecks. The next most critical set of parameters were those that determine the rate of direction choice adaptation followed by those that determine the effect of inter-individual interactions on direction choices. Conclusions: These results suggest that, in order to produce realistic evacuation times using simulation, the most critical aspect is, by far, perfecting the simulated flowrates at bottlenecks. Implications: This highlights the crucial need for furthering the accuracy of existing empirical estimates for exit capacity; and developing more rigorous calibration methods for parameters that determine flowrates at bottlenecks. If a model does not produce bottleneck flowrates accurately, efforts to refine other aspects of simulation might be in vain.