Classification-driven model selection approach of genetic programming in modelling of turning process

Turning is a widely used machining process, but the process complexity and uncertainty lead to empirical modelling techniques being preferred over physics-based models for predicting the process performance. The literature reveals that empirical methods such as artificial neural networks (ANN), support vector regression (SVR), regression analysis and fuzzy logic have been extensively applied in the modelling of turning process. The present work introduces genetic programming (GP) for the modelling of turning, but it is observed that the optimal models selected from the GP population based on training and validation errors do not perform well on testing data (unseen samples). Selecting the best GP model from the population of models is therefore a vital step. In view of this, the classification-driven model selection approach of GP (C-GP) is proposed in this paper. In this methodology, potential classification techniques such as Bayes multinomial, partitioning and regression trees, classification and regression trees and decision trees are integrated with GP to predict the class (best or bad) of the GP models. The model that is classified as the "best" by the most number of classification techniques is selected, and its performance is compared to those from ANN and SVR. It is found that the C-GP model has accuracy on par with ANN and gives satisfactory performance on testing data.