AbstractThe predictability of modulus of elasticity (MOE), modulus of rupture (MOR) and density of 120 samples of Scots pine (Pinus sylvestrisL.) were investigated using various non-destructive variables (such as time of flight of stress wave, natural frequency of longitudinal vibration, penetration depth, pullout resistance, visual grading and concentrated knot diameter ratio), and based on multivariate algorithms, applying WEKA as machine learning software. The algorithms used were: multivariate linear regression (MLR), Gaussian, Lazy, artificial neural network (ANN), Rules and decision Tree. The models were quantified based on the root-mean-square error (RMSE) and the coefficient of determination (R2). To avoid model overfitting, the modeling was built and the results validated via the so-called 10-fold cross-validation. MLR with the “greedy method” for variable selection based on the Akaike information metric (MLRak) significantly reduced the RMSE of MOR and MOE compared to univariate linear regressions (ULR). However, this reduction was not significant for density prediction. The predictability of MLRak was not improved by any other of the tested algorithms. Specifically, non-linear models, such as multilayer perceptron, did not contribute any significant improvements over linear models. Finally, MLRak models were simplified by discarding the variables that produce the lowest RMSE increment. The resulted models could be even further simplified without significant RMSE increment.
