• 목록

[Abstract]

Upon establishing multiple response optimization (MRO) or dual response optimization (DRO), response surface methodology (RSM) acts as a conventional model-driven framework for optimizing small-sized experimental data to derive an input condition in which a decent response is expected. RSM conducts multiple experiments to estimate response surfaces via empirical prediction models, but it suffers from predictive uncertainty upon fitting strict models onto large, complex operational data. Data-driven approaches on MRO and DRO have gained attention by employing supervised data mining algorithms to discover outstanding input subrange options from operational data, while effectively circumventing the predictive capability issues of RSM. In this study, we propose two new stochastic data-driven methods for response optimization to automatically derive numerous gratifying input conditions from industrial operational data. The random subspace method is incorporated into the patient rule induction method in accordance with valid objective function designs for MRO and DRO. Comparative case studies demonstrate the superior performance of our stochastic methods over deterministic baselines through N-fold validation and sensitivity analyses.