Improved predictive modeling of wood composite products using Bayesian Additive Regression Trees (BART)
Young, T. M., N. Tian, and Y. Sun.  2018.  5th Stochastic Modeling Techniques & Data Analysis (SMTDA) International Conference and Demographics.  (invited)

Abstract:
This study presents real-time prediction models of modulus of rupture (MOR) and tensile strength from a wood composites manufacturing process. Variable preselection was used in the development of the various predictive models. Several regression models including multiple linear regression, partial least squares regression, neural networks, regression trees, boosted trees, bootstrap forest, and Bayesian additive regression trees (BART) were developed. BART had the best prediction performance in validation for both MOR and tensile strength. Specifically, validation results of MOR were promising with an average correlation coefficient across cross-validations of 0.86 and prediction error of 11.8%. Validation results for tensile strength had an average correlation coefficient across validations of 0.84 and 10%. The high prediction ability of BART is useful for manufacturers and researchers in improving the manufacturing process and reducing rework and reject losses; ultimately improving business competitiveness.