33. Bringing Light to the Threshold: Evaluation of Production Policies using the Regression Discontinuity Design with Application to LED Manufacturing
Invited abstract in session WB-12: AI and Optimization for Warehousing, stream Artificial Intelligence, Machine Learning and Optimization.
Wednesday, 10:45-12:15Room: H10
Authors (first author is the speaker)
| 1. | Oliver Schacht
|
| 2. | Philipp Schwarz
|
| AMS Osram | |
| 3. | Sven Klaaßen
|
| Lehrstuhl für Statistik, Universität Hamburg | |
| 4. | Martin Spindler
|
| University of Hamburg |
Abstract
We investigate the impact of an inline rework step in a sequential lot-based manufacturing system. In this scenario, lots failing to meet specific quality thresholds, defined by multiple measurements, undergo rework to enhance quality and yield. As these thresholds are defined by a fixed value, it is challenging to compare lots with identical properties but different rework decisions. Thus, applying standard causal machine learning frameworks for estimating the rework effect and policy evaluation is difficult. We apply a framework featuring the regression discontinuity design (RDD) known from the econometrics literature to estimate the average treatment effect of the rework step locally at the cutoff. We employ recently developed approaches to improve estimation efficiency using machine learning algorithms in multi-dimensional RDD and provide insights into the expected effect of threshold adjustments on overall production yield. Additionally, we present effect identification results for multi-score RDD settings, commonly encountered in complex decision-making systems. Our simulated study using a digital twin offers practical insights for applying RDD in industrial data contexts. Finally, we apply our RDD framework to real-world data from opto-electronic semiconductor manufacturing at AMS Osram, featuring multiple RDD score dimensions, high-dimensional covariates, and fuzziness introduced by non-conforming units.
Keywords
- Machine Learning
- Statistics with Big Data
- Industrial Optimization
Status: accepted
Back to the list of papers