GPEvac: GNN-Based PPO for Adaptive Evacuation Routing During Shooting Events

AI summary

Overview

• Research area: Deep reinforcement learning, graph neural networks, and emergency evacuation planning / public safety AI.
• Technical level: Advanced. The paper assumes familiarity with graph neural networks, message passing, actor-critic reinforcement learning (PPO), and permutation-invariant architectures, though the problem framing is accessible to a general technical reader.
• Scope: The paper presents GPEvac, a GNN-based PPO framework that computes adaptive, real-time evacuation routes during active-shooter events, trained as a single policy that generalizes across different building layouts and runs in under 15 ms on CPU hardware.

What This Paper Is About

Mass shootings in public buildings have risen sharply, yet the standard guidance (the Department of Homeland Security's "Run–Hide–Fight" protocol) relies on uninformed, decentralized decisions by each individual, which produces congestion and accidental encounters with the shooter. Existing computational approaches either solve optimal routing exactly but become intractable at scale, or scale well but are rigidly tied to a single building layout and assume predictable threat behavior. This paper builds a system that reads a live graph representation of a building, tracks the shooter and crowd density, and outputs globally coordinated evacuation routes that minimize both time spent near the threat and total evacuation time.

Key Contributions

1. A topology-general evacuation policy. GPEvac combines an edge-centric graph neural network with Proximal Policy Optimization and a permutation-invariant edge-scoring actor, so a single trained network produces routing decisions for buildings of different sizes and structures rather than one policy per floor plan.

2. A message-passing scheme for long-range spatial reasoning. The architecture uses an edge-first sequential update (edge embeddings updated before node embeddings) plus a learnable virtual global node connected to every physical node, collapsing the graph diameter to at most two hops. This lets distant threats and exits inform every node's decision without the oversmoothing that comes from stacking many layers.

3. A novel reward function that couples threat avoidance with evacuation progress. The dominant term penalizes occupancy near the shooter using an exponential distance-decay function, complemented by escape bonuses, a full-evacuation terminal bonus, and a time penalty that scales with the number of people still inside.

4. Demonstrated real-time feasibility on commodity hardware. Global routing updates are computed in a median of 14.73 ms on a single Apple M4 Pro CPU, fast enough to integrate with standard surveillance camera frame rates, and inference scales sub-linearly with graph size.

Main Findings

• GPEvac beats strong heuristics on safety metrics. Averaged across both layouts, the best GPEvac policy achieved an exposure time of 0.427 versus 0.518 for the rule-based baseline and 1.274 for the greedy baseline, and a threat penalty of 4.36 versus 5.19 and 11.02 respectively.

• The improvements are statistically significant. A one-sided Wilcoxon signed-rank test (N = 64) found significant gains over the rule-based baseline in exposure time (p = 3.65 × 10⁻²), threat penalty (p = 6.71 × 10⁻⁵), and total return (p = 4.92 × 10⁻⁶).

• Safety comes at a modest speed cost. The threat-oblivious greedy policy evacuates fastest (83.75 steps on the acyclic layout) but exposes people far more. GPEvac's best policy on the cyclic layout took longer to finish (158.5 steps versus 145.97 for rule-based) while cutting exposure by roughly 25 percent, reflecting an intentional trade-off.

• The learned policy makes non-intuitive moves. Rollout analysis showed GPEvac sometimes sends evacuees who are two to three rooms from the shooter into the hallway to begin evacuating, rather than hiding, preventing them from being trapped by the shooter's later random-walk movements.

• Every architectural component matters. Ablating the edge-update mechanism raised the average threat penalty from 4.43 to 5.46 (p = 0.0186) with much higher variance; removing the virtual global node degraded performance further; removing both gave 5.63. Replacing the GNN entirely with a two-layer MLP worsened the penalty to 5.07, confirming that explicit topological encoding is necessary.

• PNA is the most reliable aggregator. Among GNN variants tested (ResGated, GATv2, Graph Transformer, GINE), Principal Neighbourhood Aggregation performed best, with GINE proving highly unstable (threat penalty 9.07). The authors attribute PNA's advantage to combining multiple aggregators that capture distinct statistics such as average risk and minimum exit distance.

• Inference scales sub-linearly. Single-step inference ranged from 11.5 ms at 50 nodes to 126.2 ms at 1,600 nodes, while one-time environment initialization grew super-linearly (0.03 s to 17.71 s) and is purely an offline preprocessing cost.

Methodology in Plain English

The researchers treat a building floor plan as a graph: each room (and discretized hallway segment) is a node, and connections between them are edges. Nodes carry features like occupant count and distance to the shooter and exits; edges carry traversal time and occupancy gradients. To make the approach practical without live video, they built a simulator that generates realistic crowd distributions (matching national averages of about 18 students per occupied classroom) and moves the shooter as a uniform random walk between adjacent nodes, deliberately avoiding assumptions about attacker behavior.

A graph neural network then processes this graph over several message-passing rounds. The key design choice is to update edge representations first, using each edge's two endpoint nodes, and then update nodes using those refreshed edges. A virtual global node wired to every physical node gives all rooms cheap access to distant information. Layer normalization and residual connections keep training stable, and padding plus masking lets the same network handle graphs of different sizes.

The resulting node and edge embeddings feed two networks. A critic estimates the overall state value by averaging all node embeddings and concatenating the global node's embedding. An actor scores every possible transition by feeding each edge's source embedding, edge embedding, and target embedding through a small MLP, producing a scalar per edge that a masked softmax converts into a routing distribution over valid neighbors. Because the actor evaluates edges individually rather than emitting a fixed-size action vector, it is unaffected by node ordering and works on any topology.

Training used PPO across 48 parallel simulated environments per layout, with both an acyclic school (strict bottlenecks) and a cyclic school (multiple escape routes) trained under a shared policy. Evaluation compared against two carefully tuned baselines: an oblivious Dijkstra shortest-path heuristic, and a rule-based policy implementing a threat-aware version of "Run–Hide–Fight" with a grid-searched safety threshold.

Why This Matters

Research impact. The paper sits at the intersection of graph representation learning and safety-critical reinforcement learning, and it offers a concrete alternative to the dominant multi-agent formulation for evacuation problems. By treating routing as a single-agent problem with a permutation-invariant edge actor, it sidesteps the non-stationarity and sample-complexity problems that the authors argue block multi-agent deployment in safety-critical settings. It also provides a template for graph-structured decision-making well beyond evacuation.

Real-world applications:

• Live campus and workplace safety systems: routed instructions pushed to digital displays and dynamic exit signage, driven by existing camera infrastructure.
• Architectural design evaluation: simulating evacuation on proposed floor plans to produce a quantitative survivability metric alongside fire-code compliance.
• Critical infrastructure planning: stadiums, transit hubs, hospitals, and industrial facilities where large crowds face dynamic hazards.
• Adjacent routing domains: intelligent transportation systems, adaptive sensor networks, and resource allocation, which share the same graph-structured decision problem.

Industry relevance. The 14.73 ms CPU inference means no GPU dependency, which is a serious practical constraint for deployment inside existing building security systems. The paper also closes a gap that has kept prior exact-optimization methods (stochastic dynamic programming, ASTERS, C-CASTERS) out of production: those methods are calibrated to specific layouts and do not scale. GPEvac's single-policy design means one trained model can serve a portfolio of buildings.

Future Directions

• Zero-shot generalization across building types. The model was trained on only two school floor plans. Compiling a large-scale architectural dataset would let researchers rigorously measure how well an unseen layout can be handled without retraining.

• More realistic threat and crowd dynamics. The current shooter is a uniform random walk. Adding capacity constraints, an adaptive adversary that reacts to evacuation flow, and stress-induced human behavior such as herding and panic would test the policy under far more demanding conditions.

• Robustness to imperfect perception. Real tracking systems make localization errors, especially in chaotic emergencies. Replacing deterministic node features with probabilistic ones would let the policy reason explicitly about observational uncertainty.

• Closing the simulation-to-reality gap. Validation inside high-fidelity platforms such as Pedestrian Dynamics or Unreal Engine is a stated prerequisite before any real deployment. An open question the paper does not resolve is how evacuees would actually comply with system-issued routing instructions under stress, and what happens when they do not.

Target Audience

This paper is most valuable to reinforcement learning and graph neural network researchers interested in safety-critical or real-time applications, and to applied researchers working on evacuation modeling, crowd dynamics, and emergency response systems. It will also interest building safety engineers, campus security administrators, and urban planners who want a quantitative tool for evaluating layout survivability, as well as industry practitioners evaluating whether learned routing policies can run within existing surveillance and building-management infrastructure.

Authors’ abstract

The sharp increase in mass shootings underscores an urgent need for systems that guide victims to safety in real time. An effective evacuation system must minimize threat exposure while also accounting for adversarial uncertainty and crowding dynamics. Current methods in the literature are rigidly constrained to layout-specific policies and computationally intractable in large-scale layouts, while practical guidelines simply advise victims to "run", "hide", or "fight". We propose GPEvac: a GNN-based PPO framework that computes adaptive evacuation routes during shooting events. To capture both local and long-distance dependencies, we introduce an edge-first sequential message-passing scheme with a learnable virtual global node. The resulting graph embeddings are integrated into a permutation-invariant scoring mechanism that allows a single learned policy to operate across building layouts of diverse topologies and sizes. Through extensive simulation, we show that GPEvac outperforms intelligent baselines across distinct architectural layouts, significantly reducing total threat exposure. Crucially, the system computes global evacuation routes in just 14.73 ms on local CPU hardware, enabling seamless integration with live surveillance systems. In addition to saving lives during shooting events, the methodologies developed are transferable to other graph-structured decision-making domains, including critical infrastructure, intelligent transportation systems, and adaptive sensor networks.