Analysis of the use of behavioral data from virtual reality for calibration of agent-based evacuation models

Agent-based evacuation modeling represents an effective tool for making predictions about evacuation aspects of buildings such as evacuation times, congestions, and maximum safe building capacity. Collection of real behavioral data for calibrating agent-based evacuation models is time-consuming, costly, and completely impossible in the case of buildings in the design phase, where predictions about evacuation behavior are especially needed. In recent years evacuation experiments conducted in virtual reality (VR) have been frequently proposed in the literature as an effective tool for collecting data about human behavior. However, empirical studies which would assess validity of VR-based data for such purposes are still rare and considerably lacking in the agent-based evacuation modeling domain. This study explores opportunities that the VR behavioral data may bring for refining outputs of agent evacuation models. To this end, this study employed multiple input settings of agent-based evacuation models (ABEMs), including those based on the data gathered from the VR evacuation experiment that mapped out evacuation behaviors of individuals within the building. Calibration and evaluation of models was based on empirical data gathered from an original evacuation exercise conducted in a real building (N = 35) and its virtual twin (N = 38). This study found that the resulting predictions of single agent models using data collected in the VR environment after proposed corrections have the potential to better predict real-world evacuation behavior while offering desirable variance in the data outputs necessary for practical applications.