• 목록

 

In multiple response optimization (MRO), data-driven approaches are often preferred over model-driven approaches in that they flexibly discover decent input conditions while circumventing predictive capability issues of model-driven approaches. Model-driven approaches for MRO attempt to construct a parametric model (i.e. second-order linear regression, RSM), but often fails due to strict model assumptions for explaining complex truth underlying large-scaled operational datasets. A non-parametric method is a representative example of data-driven approach, employing data mining algorithms on response optimization tasks. Existing non-parametric methods for MRO employ data mining methods such as classification and regression tree (CART) or patient rule induction method (PRIM). However, investing a relatively insufficient amount of algorithmic effort upon optimizing large-scaled and complex operational data show limitations upon discovering vast candidates of desirable input conditions. To overcome such limitations, we propose multiresponse random forest (MR-RF), a non-parametric data mining method for MRO which performs intensive algorithmic effort for searching desirable input variable conditions via numerous distinct tree trials. Through two comprehensive case studies, our proposed MR-RF consistently outperforms other data-driven methods by finding numerous candidates of highly desirable input conditions. Robustness of MR-RF is demonstrated through 5-fold cross-validation and sensitivity analysis.