Learning Event-Based Shooter Models from Virtual Reality Experiments

TL;DR

This work tackles the scalability gap in evaluating school-security interventions by building a data-driven discrete-event simulator (DES) trained on VR shooter experiments. The DES comprises three components: a Graph Neural Network–based shooter-transition model, a moment-matched hierarchical sampler for shooter events, and a robot-effects module driven by a smoke-diffusion kernel, enabling large-scale policy learning with minimal human-subject trials. The authors demonstrate fidelity to VR data and some external real-shooter trajectories, and they show how the DES supports rapid policy iteration, including reinforcement-learning–based control of mobile robots that reduce casualties. Limitations include reliance on a single environment and the need for VR-based validation of learned policies; nevertheless, the framework enables a principled high-to-mid fidelity workflow for autonomous mitigation in safety-critical settings.

Abstract

Virtual reality (VR) has emerged as a powerful tool for evaluating school security measures in high-risk scenarios such as school shootings, offering experimental control and high behavioral fidelity. However, assessing new interventions in VR requires recruiting new participant cohorts for each condition, making large-scale or iterative evaluation difficult. These limitations are especially restrictive when attempting to learn effective intervention strategies, which typically require many training episodes. To address this challenge, we develop a data-driven discrete-event simulator (DES) that models shooter movement and in-region actions as stochastic processes learned from participant behavior in VR studies. We use the simulator to examine the impact of a robot-based shooter intervention strategy. Once shown to reproduce key empirical patterns, the DES enables scalable evaluation and learning of intervention strategies that are infeasible to train directly with human subjects. Overall, this work demonstrates a high-to-mid fidelity simulation workflow that provides a scalable surrogate for developing and evaluating autonomous school-security interventions.

Figures (4)

• Figure 1: The annotated floor plan (left) was transformed into a graph representation (right), with regions represented as nodes and edges defining their adjacencies. Shooter behavior was modeled as a sequence of events, each defined by the time spent within a region.
• Figure 2: Greedy forward-selection results. Blue markers show the feature added at each stage of the selection process, with mean prediction accuracy and standard deviation across five folds. Black points represent the accuracies obtained by all other intermediate combinations evaluated during the search.
• Figure 3: Flowchart of the hierarchical truncated-normal sampling procedure. Event outcomes are sampled from a truncated normal distribution parameterized by region-, group-, or global-level statistical moments, depending on data availability.
• Figure 4: Prediction accuracy of region-transition models evaluated on participant data (left) and shooter data (right). Bars indicate mean accuracy with 95% confidence intervals. Bold labels above each bar report the absolute difference relative to the GNN model in percentage points. Asterisks denote significance from Welch’s unequal-variance $t$-test: *$p < 0.05$, **$p < 0.01$, ***$p < 0.001$.