Deep Bayesian Inverse Reinforcement Learning Technique to Assess Building Safety under an Active Shooter Violence Situation

Given the alarming increase in active shooter situations in the United States, improving public safety by systematically evaluating building structures has become a top priority. In particular, the majority of previous incidents took place in public, multi-storied buildings, such as shopping centers, educational institutions, and places of worship. This trend was the main motivation behind the current study to develop a cutting-edge Deep Bayesian Inverse Reinforcement Learning method to enhance our comprehension of civilians' and shooters' dynamics in sophisticated multi-storied buildings during active shooter violence incidents. We designed a grid-based educational environment where each civilian acts as an autonomous agent, trying to seek safety positions (e.g., hiding places and entrances/exits). At the same time, their actions are guided by the proposed method. Moreover, to realistically simulate multiple active shooter scenarios, this study investigates the interplay between building configurations (e.g., location and number of entrances/exits, stairs, floors, and possible hiding places) and shooter attributes (e.g., number of shooters, firearm types), to analyze their impacts on the scenario outcomes. Each scenario is evaluated under different performance metrics, such as the percentage of civilians who managed to hide, exited through entrances/exits, or are still in danger.