Modeling using Bayesian Additive Regression Trees (BART) for Wood Composite Products
Tian N., T. M. Young, A. Petutschnigg, Y. Sun, and Z. Pei.  2018.  Research Inventy: International Journal of Engineering and Science (IJSE), 7(5):66-75.

Abstract:
This study presents an ensemble of predictive models with a focus on the predictive capabilities of Bayesian Additive Regression Trees (BART). Predictions are made for Modulus of Rupture (MOR) and Tensile Strength (IB or Internal Bond) from a wood composites manufacturing process for three product types. Given the large number of predictor variables from the process, variable preselection was used prior to model development. Several regression methods including multiple linear regression, partial least squares regression, neural networks, regression trees, boosted trees, and bootstrap forest are compared with BART.BART had the best predictive performance in validation unanimously for both MOR and IB for all three products examined. Bootstrap forest validation results were very similar to BART for one of the products. BART validation results of MOR were promising for the nominal product type of 19.05 mm with an r=0.89 for 10-fold cross validation with root mean square error of prediction (NRMSEP) of 10.26%. BART validation results for IB had an average r=0.84 for10-fold cross-validation with a NRMSEP = 10.82%. The high predictive ability of BART may be useful for manufacturers and researchers in applying analytical techniques for process improvement leading to less rework (order reruns due to failing properties) and reject. Predictive modeling techniques like the ones explored in this study may be very important to companies seeking competitive advantage in today’s business world that is focused on advanced analytics and data mining.