Aegis: A cloud-edge computing based multi-disaster crowd evacuation model using improved deep reinforcement learning

Abstract

Crowd evacuation is an important measure for urban disaster management, which can provide effective evacuation guidelines for victims and safeguard their lives. However, most of existing methods are designed for single-disaster scenarios, ignoring the fact that disasters often erupt in multiple locations simultaneously. Thus, a multi-disaster crowd evacuation model, Aegis, is proposed based on cloud-edge computing and improved deep reinforcement learning. Firstly, the multi-disaster crowd evacuation problem is modeled as a multi-objective optimization problem, which considers shelter load balancing and dangerous area crossing issues. Secondly, an improved deep reinforcement learning model is proposed in this paper to solve it. The model utilizes Attention mechanism, Gated Recurrent Unit (GRU) and Graph Attention Network (GAT) to obtain the embedding of raw data. Then, the model maps the embedded information to the evacuation plan by an attention-based decoder. The model parameters are optimized using a Policy Gradient method. Thirdly, a cloud-edge computing framework is also introduced for Aegis, featuring a three-tier architecture that includes cloud, edge, and terminal levels. This design allows for the seamless integration of the model into smart city management. The experimental results show that Aegis outperforms other baseline methods, especially in reducing evacuation costs and optimizing shelter loads. In experiments with four different scales, Aegis reduces the evacuation costs by 58.87 %, 64.56 %, 65.59 %, and 67.79 %.