Uncertainty-Aware Hybrid Ensemble Learning Models for Post-Disaster Routing

Abstract: Post-disaster routing is crucial for timely relief delivery, medical aid, rescue, evacuation services, and utility restoration. Routing in post-disaster scenarios is complex due to uncertainty from road network damage and unpredictable resource demands. Factors like physical obstacles, logistical complexities, operational coordination issues, temporal constraints, and ethical considerations necessitate the development of adaptive routing strategies. These challenges require constant reassessment and adaptation of routing plans. The study proposes a post-disaster routing solution using a stochastic graph representation of road networks, incorporating probabilistic passability and demand variability. It employs a hybrid ensemble methodology, including Genetic Algorithms, Ant Colony Optimization, Particle Swarm Optimization, and Tabu Search, to enhance computational efficiency and resource allocation. The proposed ensemble approach significantly reduces travel costs (distance), outperforming benchmark algorithms across test networks. This is crucial in disaster response logistics, as even small reductions can lead to faster response times and more efficient resource distribution. This research proposes an innovative ensemble learning framework tailored explicitly for disaster relief VRPs under node-edge uncertainty. It combines a portfolio of diverse base policies, each specialized in a different uncertainty scenario, into a robust meta-policy using a stacked generalization. The Progressive Hedging Algorithm is incorporated for scenario-based planning, enhancing the model's robustness and adaptability to evolving disaster conditions. Its practical application extends to disaster response organizations, potentially improving the overall effectiveness and outcomes of aid delivery in areas stricken by disasters.
The framework dynamically adapts routing decisions to emerging needs and infrastructure damage by blending complementary strengths. This research contributes to the existing literature by developing an innovative ensemble learning framework tailored for stochastic VRPs in post-disaster contexts. The proposed model is expected to substantially improve relief efficiency and survival rates after disasters compared to current VRP methods.