Abstract The behaviour of building occupants in the first stage of an evacuation can dramatically impact the time required to evacuate buildings. This behaviour has been widely investigated by scholars with a macroscopic approach fitting random distributions to represent the pre-evacuation time, i.e. time from noticing the first cue until deliberate movement. However, microscopic investigations on how building occupants respond to several social and environmental factors are still rare in the literature. This paper aims to leverage machine learning as a possible solution to investigate factors affecting building occupants' decision-making during pre-evacuation stage. In particular, we focus on applying interpretable machine learning to reveal the interactions among the input variables and to capture nonlinear relationships between the input variables and the outcome. As such, we use a well-established machine-learning algorithm—random forest—to model and predict people's emergency behaviour pre-evacuation. We then apply tools to interpret the black-box random forest model to extract useful knowledge and gain insights for emergency planning. Specifically, this algorithm is applied here to investigate the behaviour of 569 building occupants split between five unannounced evacuation drills in a cinema theatre. The results indicate that both social and environmental factors affect the probability of responding. Several independent variables, such as the time elapsed after the alarm has started and the decision-maker's group size, are presenting strong nonlinear relationships with the probability of switching to the response stage. Furthermore, we find interactions exist between the row number where the decision-maker sits and the number of responding occupants visible to her; the complex relationship between the outcome and these two variables can be visualized by using a two-dimensional partial dependence plot. An interesting finding is that a decision-maker is more sensitive to the proportion of responding occupants than the number of them; hence, the people sitting in the back are often responding more slowly than the people in the front.

中文翻译：

---

使用机器学习对疏散前的决策进行建模和解释

摘要 建筑物居民在疏散第一阶段的行为会极大地影响疏散建筑物所需的时间。这种行为已被学者广泛研究，采用宏观方法拟合随机分布来表示预疏散时间，即从注意到第一个线索到有意移动的时间。然而，关于建筑物居住者如何对几种社会和环境因素做出反应的微观研究在文献中仍然很少见。本文旨在利用机器学习作为一种可能的解决方案来调查影响建筑物居住者在预疏散阶段决策的因素。特别是，我们专注于应用可解释的机器学习来揭示输入变量之间的相互作用并捕获输入变量与结果之间的非线性关系。因此，我们使用成熟的机器学习算法——随机森林——来模拟和预测人们在疏散前的紧急行为。然后我们应用工具来解释黑盒随机森林模型，以提取有用的知识并获得应急计划的见解。具体来说，这里应用该算法来调查 569 名建筑物居住者的行为，这些人员在电影院进行了五次突击疏散演习。结果表明，社会和环境因素都会影响响应的概率。几个自变量，例如警报开始后经过的时间和决策者的 s 组大小，与切换到响应阶段的概率呈现出很强的非线性关系。此外，我们发现决策者所在的行号与她可见的响应乘员数量之间存在交互作用；结果与这两个变量之间的复杂关系可以通过使用二维部分依赖图来可视化。一个有趣的发现是，决策者对做出反应的住户比例比对住户人数更敏感；因此，坐在后面的人往往比前面的人反应慢。我们发现决策者所在的行号与她可见的响应乘员数量之间存在交互作用；结果与这两个变量之间的复杂关系可以通过使用二维部分依赖图来可视化。一个有趣的发现是，决策者对做出反应的居住者的比例比他们的人数更敏感；因此，坐在后面的人往往比前面的人反应慢。我们发现决策者所在的行号与她可见的响应乘员数量之间存在交互作用；结果与这两个变量之间的复杂关系可以通过使用二维部分依赖图来可视化。一个有趣的发现是，决策者对做出反应的居住者的比例比他们的人数更敏感；因此，坐在后面的人往往比前面的人反应慢。