GPEvac
A real-time evacuation path planning system combining graph neural networks (GNN) with the PPO reinforcement learning algorithm. It achieves adaptive evacuation route computation across diverse building layouts through an edge-prioritized message-passing mechanism and a learnable virtual global node.
Timeline (last 90 days)
GPEvac is a system that combines Graph Neural Networks (GNN) with the PPO (Proximal Policy Optimization) algorithm to compute adaptive escape routes in real time during shooting incidents.
GPEvac introduces an edge-first sequential message-passing mechanism and a learnable virtual global node.
GPEvac adopts a permutation-invariant scoring mechanism, allowing a single trained policy to be directly transferred to building layouts with varying topologies and scales without retraining.
In simulation validation, GPEvac outperforms intelligent baseline methods across different building layouts and can significantly reduce the total threat exposure of victims.
GPEvac requires only 14.73 milliseconds to compute global evacuation routes on local CPU hardware.
The GPEvac research results currently remain in a simulated environment and have not yet been validated in real shooting scenarios.
The research team points out that existing evacuation path methods have two major problems: over-reliance on specific building layouts and excessively high computational complexity.
The virtual global node is directly connected to all real nodes, allowing information exchange between any two points to be completed in at most two hops, solving the problem of long-distance dependency attenuation.
The research team believes that GPEvac's methodology can be transferred to graph-structured dynamic decision problems such as critical infrastructure scheduling, intelligent traffic route optimization, and adaptive sensor network resource allocation.
All Facts (9)
GPEvac is a system that combines Graph Neural Networks (GNN) with the PPO (Proximal Policy Optimization) algorithm to compute adaptive escape routes in real time during shooting incidents.
50%UnverifiedGPEvac introduces an edge-first sequential message-passing mechanism and a learnable virtual global node.
50%UnverifiedGPEvac adopts a permutation-invariant scoring mechanism, allowing a single trained policy to be directly transferred to building layouts with varying topologies and scales without retraining.
50%UnverifiedIn simulation validation, GPEvac outperforms intelligent baseline methods across different building layouts and can significantly reduce the total threat exposure of victims.
50%UnverifiedGPEvac requires only 14.73 milliseconds to compute global evacuation routes on local CPU hardware.
50%UnverifiedThe GPEvac research results currently remain in a simulated environment and have not yet been validated in real shooting scenarios.
50%UnverifiedThe research team points out that existing evacuation path methods have two major problems: over-reliance on specific building layouts and excessively high computational complexity.
50%UnverifiedThe virtual global node is directly connected to all real nodes, allowing information exchange between any two points to be completed in at most two hops, solving the problem of long-distance dependency attenuation.
50%UnverifiedThe research team believes that GPEvac's methodology can be transferred to graph-structured dynamic decision problems such as critical infrastructure scheduling, intelligent traffic route optimization, and adaptive sensor network resource allocation.
50%