LOGEX: Cost-Sensitive Bayesian Experimentation for Adaptive Decision-Making in Supply Chains

Lorenzo Toni

Keywords: Bayesian optimization, supply chain experimentation, value of experimentation, BART, partial rollout, non-stationary rewards, cost-sensitive learning

Abstract:

Running real-world experiments in supply chains is costly, risky, and often limited by operational constraints. Evaluating a new policy—such as a revised inventory heuristic or a routing strategy—requires partial deployment, active monitoring, and foregone opportunity from not exploiting the current best-known alternative. To address this, we propose LOGEX, a Bayesian framework for cost-sensitive experimentation that models uncertain, evolving reward functions using Bayesian Additive Regression Trees (BART). LOGEX quantifies the expected value of experimentation in economic units, enabling practitioners to weigh the benefit of learning against the cost of conducting operational pilots. Unlike conventional black-box optimization, our approach supports partial rollouts, adapts to non-stationary reward landscapes, and maintains interpretability through rule-based posterior estimates. We validate LOGEX in a synthetic supply chain environment and show that it outperforms cost-unaware exploration strategies, achieving higher cumulative reward with fewer, more valuable experiments. The framework offers a practical and theoretically grounded solution for high-stakes experimentation in logistics and operations.

This work represents an ongoing collaboration between Amazon's Supply Chain Execution team and the broader research community. We present LOGEX as both a practical solution for industrial-scale supply chain experimentation and a theoretical contribution to cost-aware Bayesian optimization. Our goal is to share methodological insights with the KDD community while highlighting the substantial impact that principled experimentation frameworks can achieve in real-world operational settings where decisions affect global supply chain performance.
The framework demonstrates how academic advances in Bayesian optimization can be successfully adapted for production deployment in complex, high-stakes operational environments. We seek community feedback on our approach and aim to establish new benchmarks for evaluating experimentation frameworks in supply chain contexts where traditional metrics fail to capture the full economic complexity of operational decision-making.

Submission Number: 25