Optimizing parameter design: the univariate forest-genetic method

Abstract

In the 80’s, Dr Genichi Taguchi developed a methodology for processes and product parameters design improvement known as the Taguchi methodology. Different proposals have emerged involving artificial intelligence techniques. Our proposal consists of a hybrid methodology that combines Random Forest (RF) and Genetic Algorithms (GA) in three phases: normalization, modeling and optimization. The first phase corresponds to the previous preparation of the data set by using normalization functions. In the modeling, the objective function is determined using strategies based on RF to predict the value of the response in a given set of parameters. Finally, in the optimization phase, the optimal combination of the parameter levels is obtained by integrating properties given by our modeling scheme into the corresponding GA. The results are compared numerically with the contributions recently found in the literature. Our methodological proposal focuses on the most important variables resulting from the RF modeling process, which allows to develop and direct more efficiently the new generations in the optimization phase, and consequently, achieve significant improvements in the quality objective considered.