Regularization and concave loss functions for estimation of chemical kinetic models

Opara, K. & Oh, P.P. (2022). Regularization and concave loss functions for estimation of chemical kinetic models. Applied Soft Computing 116 e108286. 10.1016/j.asoc.2021.108286.

Full text not available from this repository.

Abstract

Non-linear regression is the primary tool for estimating kinetic models of chemical reactions. The default approach of minimizing the sum of squared residuals tends to underperform in the presence of systematic errors, non-normal distribution of residuals or identifiability issues such as a high correlation between parameters. Therefore, we argue for a careful choice of the fit criteria and propose new, concave loss functions. Together with regularization, they form a robust objective for the regression procedure. Discussion of the rationale behind the proposed approach and its effects is illustrated by laboratory data on the transesterification of palm oil. A dedicated simulation study complements qualitative examples. All of the top-performing methods use regularization. Concave loss functions were among the best in 6–7 out of 8 test cases, compared to 2–3 for the classical square loss confirming both statistical and practical usefulness of the novel fit criteria. This result holds for a variety of modern optimizers. In 76% of our simulations, we obtained results not significantly worse than the best, whereas methods currently used in the literature provide 38% for the relative and 0% for the square loss.

Item Type:	Article
Research Programs:	Advanced Systems Analysis (ASA) Young Scientists Summer Program (YSSP)
Depositing User:	Luke Kirwan
Date Deposited:	04 Jan 2022 09:22
Last Modified:	04 Jan 2022 09:22
URI:	https://pure.iiasa.ac.at/17723

Actions (login required)

View Item