Optimized Evacuations for Active Shooter Incidents to Improve Safety Outcomes

Active shooter incidents (ASIs) represent a uniquely complex and time-critical challenge for emergency evacuation planning, where traditional protocols such as “Run, Hide, Fight” often fall short due to the dynamic nature of the threat and the cognitive overload experienced by evacuees. This dissertation presents a comprehensive framework for optimizing evacuee routing during ASIs using both model-based and model-free reinforcement learning (RL) techniques. The problem is formulated as a Non-Homogeneous Semi-Markov Decision Process (NHSMDP), capturing the temporal and spatial dynamics of both evacuees and the shooter within a graph-based representation of building layouts. Three major contributions are made: (1) the development of Naive ASTERS, a model-based RL algorithm that incorporates shooter location, node safety, and exit proximity to generate optimal routing policies; (2) the extension to C-CASTERS, which introduces capacity constraints and iteratively generates multi-route evacuation plans to reduce bottlenecks and crowding; and (3) the integration of model-free approaches—Q- Learning and N-Step Temporal Difference Learning—trained in a discrete event simulation (DES) and evaluated in a high-fidelity, physics-based Unreal Engine 5 (UE5) environment. Across thousands of simulated scenarios, N-Step Temporal Difference Learning consistently outperformed other methods in survival rate, casualty reduction, and adaptability to dynamic threats. This work advances the state of the art in emergency evacuation modeling by demonstrating the viability of offline-trained RL policies in real-time, physics-based environments. It also lays the groundwork for future integration of virtual reality training, multi-agent coordination, and real-time decision support systems, with the ultimate goal of enhancing public safety and survivability during active shooter events.