Memory based local learning: application to process engineering problems

Abstract

Linear and nonlinear regression problems are very common in different fields of science and engineering. A number of different methods to provide accurate input/output relationship have been proposed in the literature. Lazy learning is a recently introduced nonparametric regression technique that employs a memory based local learning approach. Employment of (i) local weighted regression for parametric identification and (ii) PRESS statistic to assess a local model for structural identification, are the two unique features of Lazy learning. The method attempts to fit the training data in a region around the location of the query point by locally interpolating the relevant neighboring points based on distance measure. This is particularly useful when limited number of input/output pairs are available and an accurate prediction is required. Being a memory based technique, it does not require separate training, which greatly enhances the speed of implementation. Moreover the method is less susceptible to noise contamination. All these features are illustrated with one benchmark multivariate statistical example and three important chemical engineering examples comprising boiling point prediction of aliphatic hydrocarbons using quantitative structural property relations, system identification for a continuous stirred tank reactor in which an exothermal irreversible reaction between sodium thiosulfate and hydrogen peroxide occurs and a fermentation process for polyol production. The performance of the proposed method is compared with some of the state-of-art approaches.