Proceedings of the 24th International Academic Conference, Barcelona

USING GENERALIZED PATHSEEKER REGULARIZED REGRESSION FOR MODELING AND PREDICTION OF OUTPUT POWER OF CUBR LASER

SNEZHANA GOCHEVA-ILIEVA, ILIYCHO ILIEV

Abstract:

A Generalized PathSeeker Regularized Regression (GPSRR), based on data mining approach, is applied for statistical modeling and prediction of output power of copper bromide vapor lasers. The aim is on the basis of available experimental data to construct appropriate predictive models of the output power of the lasers depending on 10 operating laser characteristics in order to direct future experiments and designing new laser devices with increased output power. In particular, the influence on model performance and predictive ability of several data transformations, used to improve the normality of the distribution of the dependent variable is investigated. As a main result, numerous combined models, built by GPSRR with data mining techniques are obtained and their adequacy is established by cross-validation. It is found that the best combined models demonstrate up to 98-99% of fitting the experimental data. The combined models with the proposed preliminary transformations improve the adequacy and predictive ability of GPSRR in the region of high values of the output power by up to 10%. This was established both for learn and test random samples, showing a perfect out-of-sample performance of this type of model approach. The models are applied for predicting of laser output power for new laser devices of the same type by up to 15%.

Keywords: Regularized regression, Generalized PathSeeker, LASSO, TreeNet (Stochastic Gradient Boosting), Copper bromide vapor laser

DOI: 10.20472/IAC.2016.024.033

PDF: Download