Speaker
Description
Model selection in screening experiments is challenging when data exhibit random effects arising from split-plot structures, batch variation, or other sources of non-independent errors. Standard approaches such as stepwise selection, LASSO, and mixed-integer optimisation (MIO) typically assume independent errors or provide limited support for the correlation structures commonly encountered in industrial experiments, restricting their applicability in practice.
We present an adaptation of the Simulated Annealing Model Search algorithm (SAMS; Wolters and Bingham, 2011) that accounts for correlation induced by random effects, enabling computationally efficient model search while retaining generalized least squares or GLS-based inference for final model selection. The adapted algorithm retains the main strengths of SAMS: exploration of a large collection of well-fitting models, visualisation of effect aliasing via raster plots, and identification of active effects using an entropy-based criterion that quantifies how consistently terms appear across the model set. The method is implemented in the open-source Python package PyOptEx.
We evaluate the approach through a simulation study covering a range of experimental designs and variance structures, including a real split-plot type of screening experiment from potato fry production. We also apply the method to the data from the potato fry experiment to demonstrate practical applicability. Results show robust recovery of active effects across a range of designs and variance structures. Additional simulations investigate the performance of SAMS under strong random effects, where methods that ignore correlation structure, including MIO and backward selection, show systematic degradation in performance.
| Keywords | Screening experiments; random effects; split-plot designs |
|---|