A Comparative Analysis of Explainability Techniques for Multi-Objective Reinforcement Learning in Humanitarian Logistics

Deep Reinforcement Learning (DRL) has emerged as a powerful approach for optimizing multi-objective humanitarian logistics, but its "black-box" nature creates critical barriers to adoption in high-stakes crisis response. While prior work has demonstrated that Pareto-aware DRL agents can significantly outperform traditional methods in balancing speed, cost, and equity in aid distribution, the lack of transparency undermines trust and accountability-core humanitarian principles. This paper addresses a critical gap: the absence of systematic comparisons of explainable AI (XAI) techniques for multi-objective DRL in humanitarian contexts. Three prominent post-hoc XAI methods-SHAP, LIME, and Integrated Gradients- applied to a validated Pareto-aware Deep Q-Network for crisis resource allocation. Analyzing ۱,۰۰۰ decision explanations across four quantitative metrics reveals fundamental trade-offs: SHAP achieves the highest faithfulness and consistency, making it ideal for strategic auditing. In contrast, LIME generates explanations ۳× faster with superior sparsity but suffers from significant instability and lower faithfulness. Integrated Gradients offers a compelling compromise with strong faithfulness, high consistency, and low computational cost, positioning it as optimal for real-time field operations. These findings provide humanitarian organizations with an evidence-based framework for selecting XAI tools tailored to specific operational contexts from post-crisis audits requiring deep fidelity to time-sensitive field decisions demanding rapid, actionable insights-ultimately enabling responsible deployment of advanced AI in life-critical environments.

کلیدواژه ها:

Explainable Reinforcement AI ، Deep Learning ، Humanitarian Logistics ، Multi- Objective Optimization ، SHAP ، LIME