A NOVEL ENSEMBLE BASE MODELS SELECTION APPROACH FOR ESTIMATING CREDIBLE RUBBER CONVEYOR BELTING CURE TIMES FROM A SMALL SAMPLE SIZE MRSM DATASET

Abstract

Multiple response surface methodology (MRSM) is small sample dataset analytics in nature. There are experiences associated with small sample size dataset problems in regression models, model selection and generalisability of models that affect the MRSM solution credibility. In this work, ensembling is used to account for the small sample size problems of an MRSM dataset and to avoid the tradition of simultaneously optimizing only single “best” models for each response. A novel ensemble base model(s) selection methodology is used to manage the number of simultaneous optimization computations. Fifteen model selection criteria are used to vote for the “best” fitting and parsimonious model and, with it, all response models nesting it are included as base models of the ensemble. Simultaneous optimizations and frequency analysis of solutions and model complexities are performed to arrive at a solution. The ensemble solution is estimated by weighted averaging of simultaneous optimization solutions using the solution frequencies. When the methodology is applied to the rubber-covered conveyor belt problem of Pavolo and Chikobvu (2022), the same estimated credible results are obtained, albeit with fewer simultaneous optimization computations. The results also show that the “best” fitting and parsimonious model is an under-fit, and its solution is different from the credible ensemble solution. The approach is recommended for similar small sample size problems.

Keywords

• Multiple response surface methodology
• Ensembling
• Small data analytics
• Simultaneous optimisation
• Model selection criteria
• Nesting