EURO 2024 Copenhagen
Abstract Submission

EURO-Online login

3019. Navigating parametric cost function approximation: A practitioner's guide for sequential decision-making with integer constraints in a production setting

Invited abstract in session WD-34: New problems in logistics under uncertainty, stream Stochastic, Robust and Distributionally Robust Optimization.

Wednesday, 14:30-16:00
Room: 43 (building: 303A)

Authors (first author is the speaker)

1. Max Barkhausen
tulanā

Abstract

Warren Powell recently coined the term "parametric cost function approximation" (CFA) to describe a practical approach for addressing multistage stochastic decision problems in a deterministic manner. Despite its widespread industrial use, Powell notes a lack of scientific rigor in assessing its effectiveness. In this study, we apply CFA to a real-world multistage lot sizing and inventory management challenge encountered in the automotive sector, where stochastic programming techniques such as SDDiP come with weak convergence guarantees and unacceptable run times. We aim to highlight two key decision points in crafting a CFA algorithm and propose scientific methodologies to tackle them. Firstly, we discuss methods for selecting appropriate parameterization techniques. Second, we explore strategies for identifying optimal parameter values, utilizing Monte Carlo simulation and derivative-free optimization methods. Moreover, we emphasize and explain the importance of evaluating not just the expected value but also the risk associated with resulting policies. Finally, we outline practical strategies for deploying CFA algorithms effectively in production settings marked by varying problem instances and a rolling planning window. Through our investigation, we contribute to enhancing the methodological soundness and practical utility of CFA in real-world decision-making contexts.

Keywords

Status: accepted

Back to the list of papers