Evolve-DGN: An Evolving Dynamic Graph Network for Adaptive and Equitable Resource Allocation in Disaster Response

Abstract

The effective distribution of resources during and after a disaster is a problem of immense complexity and critical importance. As disaster situations unfold, the network of affected areas, available resources, and viable transportation routes changes dynamically, rendering static optimization models ineffective. Existing machine learning approaches often fail to capture the complex, evolving spatio-temporal dependencies or handle the frequent topological changes inherent in a crisis zone. This paper introduces Evolve-DGN, a novel framework for adaptive and equitable emergency resource allocation. Evolve-DGN models the disaster environment as a dynamic graph and leverages a unique combination of an evolving dynamic graph neural network and multi-agent reinforcement learning (MARL). The core of the framework is a GNN architecture that evolves its parameters over time, enabling it to adapt to real-time changes in the network topology, including the appearance and disappearance of nodes and edges. This GNN serves as a powerful state encoder for a cooperative MARL system where resource depots act as decentralized agents, learning to make coordinated dispatch decisions. A key contribution is the design of a multiobjective reward function that explicitly promotes efficiency, effectiveness, and equity in resource distribution, addressing a well-documented gap between academic models and practitioner needs. The efficacy of Evolve-DGN is demonstrated in a high-fidelity simulation environment, where it consistently outperforms other learning-based baselines in minimizing resource delivery time, a critical factor in saving lives, while maintaining competitive performance in overall resource distribution.