Quantifying interactions in manufacturing using regression tree models – A useful inductive step for planning a designed experiment
Young, T. M.  2018.  29th ICWST International Conference on Wood Science and Technology, Dec. 6-7, 2018. Zagreb, Croatia.  (invited)

Quantifying interactions effects for independent variables in R&D experimentation is essential for discovery. Quantifying variation from interaction effects of parameters in industrial applications is the genesis for a higher level of improvement. One of the challenges in R&D is preselecting the factors and the levels of factors to be investigated. Regression trees (RTs) in both R&D and industrial applications can identify unknown interaction effects. RTs when applied to nonhomogeneous data may result in a smaller generalized error for Y relative to other supervised learning methods while creating a visualization of a hierarchy interaction effects. RTs were used to quantify the interaction effects for the internal bond (IB) strength of medium density fiberboard (MDF). The significant interaction effects were: ‘weight set-point;’ ‘core fiber-moisture;’ and ‘pressing time.’ These factors with main and two-level interaction effects were used to create a Box-Behnken response surface model (RSM) with two replicates and three center-points for a total of 15 experimental runs. The response surface model had nine terms and simulations revealed a maximization of IB for the interaction ‘weight set point and ‘core fiber moisture’ when ‘weight set points’ ranged between 2.9 and 3.3 in the presence of a ‘core fiber moisture’ 9.2% and 9.8%.