The sharp rise in mass‑shooting incidents creates an urgent demand for systems that can guide victims to safety in real time. An effective evacuation solution must minimize threat exposure while accounting for adversarial uncertainty and crowd dynamics. Existing approaches are typically confined to layout‑specific policies and become computationally prohibitive in large‑scale environments; practical advice often reduces to “run, hide, or fight”.
We introduce GPEvac, a framework that leverages Graph Neural Networks (GNN) together with Proximal Policy Optimization (PPO) to compute adaptive evacuation routes during shooting events. To capture both local and long‑range dependencies, we employ an edge‑first sequential message‑passing scheme and incorporate a learnable virtual global node that aggregates information across the entire graph. The resulting graph embeddings are fed into a permutation‑invariant scoring mechanism, enabling a single learned policy to operate across building layouts of varying topologies and sizes.
Extensive simulations demonstrate that GPEvac consistently reduces total threat exposure compared with intelligent baselines across diverse architectural configurations. Crucially, the system computes global evacuation routes in only 14.73 ms on a local CPU, making it suitable for seamless integration with live surveillance platforms.
Beyond shooting incidents, the proposed methodology is transferable to other graph‑structured decision‑making domains such as critical infrastructure protection, intelligent transportation systems, and adaptive sensor networks.
Review: GPEvac elegantly fuses graph neural networks with reinforcement learning to achieve fast, layout‑agnostic evacuation planning, offering a practical and scalable tool for emergency response.