Robust Multi-Criteria Decision Support for AI-Assisted Marketplace Logistics with Uncertain Data and Preferences

Timur Panesh, Anna Kovalenko

Keywords: robust optimization, multi-criteria decision making, AI-assisted decision support, minimax regret, uncertainty, marketplace logistics

TL;DR: We propose a robust multi-criteria decision support framework for AI-assisted marketplace logistics that explicitly accounts for uncertainty in both data and preference weights.

Abstract: Decisions in marketplace logistics and dropshipping involve conflicting objectives such as cost, delivery time, reliability, and returns, and are increasingly supported by AI-assisted decision support systems that combine data-driven analytics with human judgment. In practice, however, both performance data and preference weights used in multi-criteria decision-making (MCDM) are subject to significant uncertainty, arising from noisy data pipelines, reporting delays, and variability in human-in-the-loop preference elicitation. As a result, solutions derived from fixed-parameter MCDM models may be fragile and sensitive to modest perturbations. This paper develops a robust multi-criteria decision support framework that serves as a mathematical core for AI-assisted decision-making under uncertainty. Uncertainty is explicitly modeled in both the criteria matrix and the preference weights, and the decision problem is formulated using an additive utility model as a max–min optimization that maximizes guaranteed utility over a joint uncertainty set. In addition, a minimax-regret formulation is considered to bound the maximum loss relative to scenario-wise optimal decisions, which is particularly relevant for AI-enabled recommendation and decision support settings. Beyond optimal decisions, the framework provides quantitative diagnostics for AI-supported analysis, including stability regions, regret profiles, the Expected Value of Perfect Information, and the Value of Uncertainty Reduction, enabling transparent robustness assessment and guiding improvements in data quality and model calibration. A numerical illustration demonstrates that robust strategies embedded in AI-assisted systems trade off some nominal performance in favorable scenarios while substantially improving worst-case outcomes and reducing exposure to adverse conditions.

Submission Number: 19