Adaptive vehicle routing for humanitarian aid in conflict-affected regions: a practitioner-informed deep reinforcement learning approach

Humanitarian aid delivery in conflict-affected regions faces significant challenges due to dynamic security risks, uncertain demand, and complex operational constraints. Traditional optimization methods struggle with computational intractability and lack adaptability for real-time decision-making in volatile environments. To address these limitations, we propose a novel hybrid framework that integrates Deep Reinforcement Learning (DRL) with Graph Neural Networks (GNNs) and deterministic constraint validation, informed by practitioner insights to ensure real-world applicability. Our approach employs Proximal Policy Optimization (PPO) enhanced by GNN-based spatial representations to learn adaptive, efficient vehicle routing policies under uncertainty. A post-decision validation mechanism enforces feasibility by penalizing constraint violations based on a deterministic equivalent model. We evaluate our method on realistic, georeferenced datasets reflecting Afghan road networks and conflict data, comparing it against classical PPO and heuristic baselines. Results demonstrate that PPO-GNN significantly reduces operational costs (by 7.9%), security risk exposure (by 15.2%), and unmet demand, while improving reliability and adherence to constraints. The approach scales effectively across network sizes and maintains robustness under stochastic variations in demand and security conditions. Our framework balances computational efficiency with practical relevance, aligning with humanitarian priorities and offering a promising decision-support tool for aid logistics in conflict zones.