Causal effect estimation framework for early human decision-making under nuclear emergencies: Integrating virtual reality and machine learning

Abstract

Understanding the causal factors of early emergency responses, particularly in the context of nuclear emergencies, is crucial for disaster management. However, challenges arise from the specificity of disasters, uncertainty in human behavior, and the lack of reliable data. Virtual Reality (VR) and Machine Learning (ML) have emerged as promising tools for addressing these challenges. This study proposes a novel framework for estimating causal effects using VR and ML technologies. The proposed framework was designed to address the research questions identified in the research design phase and primarily consists of two key components: (1) the implementation of a VR-based Human-in-the-Loop (HITL) experimental system for extracting human behavior data, and (2) the development of a ML-based causal model utilizing the collected behavioral data. In the first component, the VR-based HITL experimental system focuses on the process of setting up disaster scenarios and collecting behavioral data in response to the proposed causal factors. Subsequently, the second component involves using the collected behavioral data to develop a ML-based causal model that quantitatively estimates causal effects by controlling for biases and confounding variables. The verification experiment results demonstrated that the nuclear emergency scenario (ATT: -94.08 sec) significantly reduced decision-making time more than any other cue. Notably, the scenario in which a clear and accurate evacuation order was provided by the government (Scenario 2, ATT: -116.70 sec) saw the most substantial reduction in decision-making time of the scenarios. Although official cues (ATE: -16.15 sec) had a relatively limited effect compared to the emergency scenario (ATT: -94.08 sec), they underscore the need to enhance government guidelines to facilitate quicker decision-making. The proposed framework, through verification experiments, illustrates how modern techniques can estimate causal relationships in decision-making during emergencies, enabling disaster management experts to make reliable data-driven decisions.