RERolf Ergon

Rolf Ergon

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Rolf Ergon has a total of 3 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    When three traits make a line: Evolution of phenotypic plasticity and genetic assimilation through linear reaction norms in stochastic environments

    by Torbjørn Ergon, Rolf Ergon

    Genetic assimilation results from selection on phenotypic plasticity, but quantitative genetics models of linear reaction norms considering intercept and slope as traits do not fully incorporate the process of genetic assimilation. We argue that intercept-slope reaction norm models are insufficient representations of genetic effects on linear reaction norms, and that considering reaction norm intercept as a trait is unfortunate because the definition of this trait relates to a specific environmental value (zero) and confounds genetic effects on reaction norm elevation with genetic effects on environmental perception. Instead we suggest a model with three traits representing genetic effects that respectively (i) are independent of the environment, (ii) alter the sensitivity of the phenotype to the environment, and (iii) determine how the organism perceives the environment. The model predicts that, given sufficient additive genetic variation in environmental perception, the environmental value at which reaction norms tend to cross will respond rapidly to selection after an abrupt environmental change, and eventually become equal to the new mean environment. This readjustment of the zone of canalization becomes completed without changes in genetic correlations, genetic drift or imposing any fitness costs on maintaining plasticity. The asymptotic evolutionary outcome of this three-trait linear reaction norm generally entails a lower degree of phenotypic plasticity than the two-trait model, and maximum expected fitness does not occur at the mean trait values in the population.

  • Model choice and squared prediction errors in PLS regression

    by Rolf Ergon, Maths Halstensen, Kim H. Esbensen

    Squared prediction errors (SPE) in are discussed in relation to the conventional PLSR versus bidiagonalization model and algorithm issue concerning residual and prediction consistency, with focus on process monitoring and fault detection. Our analysis leads to the conclusion that conventional PLSR based on the NIPALS algorithm is ambiguous in SPE values caused by process faults. The basic reason for this is that the sample residuals are not found as projections onto the orthogonal complement of the space where the scores and regression solution are located, and where also the statistical limit is defined. The alternative non‐orthogonalized PLSR and bidiagonalization (Bidiag2) algorithms, as well as a simple re‐formulation of the NIPALS algorithm (RE‐PLSR), give unambiguous SPE values, and the last two of these also retain orthogonal score vectors. While prediction results from all of these methods in theory are identical, our conclusion is that methods where the and SPE values for process faults are uncorrelated should be preferred. Tests with added errors on real data do not indicate that this conclusion should be altered because of such errors. Copyright © 2011 John Wiley & Sons, Ltd.


  • Re‐interpretation of NIPALS results solves PLSR inconsistency problem

    by R. Ergon

    The well‐known nonlinear iterative partial least squares (NIPALS) algorithm is commonly used for computation of components in partial least squares regression (PLSR) with orthogonalized score vectors. Based on generalized inverse formalism, Pell et al.[1] have recently claimed that the NIPALS results are inconsistent with respect to model spaces for residual‐based outlier detection and prediction purposes. This theoretically important result is supported by the present paper, where it is also shown that a simple re‐interpretation of the results from the NIPALS algorithm solves the problem. This is valid for cases with both one and several response variables (PLS1 and PLS2). It is also shown, however, that the price to pay for this solution is that the latent variables in the model will no longer be completely independent of the residual noise. Since the original PLSR model with non‐orthogonal score vectors does not have the same inconsistency, a method for orthogonalization of this model is also included. Copyright © 2008 John Wiley & Sons, Ltd.


  • Finding Y‐relevant part of X by use of PCR and PLSR model reduction methods

    by Rolf Ergon

    The paper is considering the following question: using principal component regression (PCR) or partial least squares regression (PLSR), how much data can be removed from X while retaining the original ability to predict Y? Two model reduction methods using similarity transformations are discussed, one giving projections of original loadings onto the column space of the fitted response matrix $\hat {\bf Y}$ (essentially the orthogonal signal correction (OSC) methods), and one giving projections of original scores onto the column space of the coefficient matrix $\hat {\bf B}$ (essentially the net analyte signal (NAS) methods). The loading projection method gives model residuals that are orthogonal to Y and $\hat {\bf Y}$ , which is valuable in certain applications. The score projection method, on the other hand, gives model residuals that are orthogonal to $\hat {\bf B}$ , which is essential in other applications. It is shown that the reduced matrix X from the score projection method is a subset of the reduced matrix X from the loading projection method. It therefore has the smallest Frobenius norm, and thus the smallest total column variance, assuming centered data. Copyright © 2007 John Wiley & Sons, Ltd.


  • PLS post‐processing by similarity transformation (PLS + ST): a simple alternative to OPLS

    by Rolf Ergon

    Several methods for orthogonal signal correction (OSC) based on pre‐processing of the modeling data have been developed in recent years, and OPLS (orthogonal projections to latent structures) is a well‐known algorithm. The main result from these methods is a reduction in the number of final components in partial least squares (PLS) regression, while the predictions are virtually unchanged (identical for OPLS). This raises the question whether the same or similar results can be obtained in a more direct way using an ordinary PLS model as starting point, and as shown in the present paper, this can indeed be done by use of a simple similarity transformation. This post‐processing PLS + ST method is compared with OPLS assuming a single response variable. The PLS + ST factorization of the data matrix X is just a similarity transformation of the non‐orthogonalized PLS factorization, while OPLS is a similarity transformation of the orthogonalized PLS factorization. The predictions are therefore identical but the residuals are somewhat different. A theoretically founded modification of the orthogonalized PLS factorization, and a corresponding modification of OPLS, leads to identical factorizations for all these methods within similarity transformations. The PLS + ST vs OPLS comparison also leads to an alternative post‐processing method using the ordinary PLS algorithm twice, with predetermined and permuted loading weight vectors in the second step. A limited comparison with post‐processing using principal components of predictions (PCP) or canonical correlation analysis (CCA) is included. Copyright © 2005 John Wiley & Sons, Ltd.


  • Compression into two‐component PLS factorizations

    by Rolf Ergon

    Partial least squares regression (PLSR) often requires more than two components also in the case of a scalar response variable. As shown in papers on orthogonal signal correction (OSC), it is possible to reduce the number of components, resulting in easier data interpretation. In this paper it is shown how all scalar response PLSR models can be reduced to two‐component models with the same structure and giving exactly the same estimator as the original model using many components. This is done by use of a direct and very simple algorithm based on a two‐dimensional subspace in the loading weight space. The resulting model may be transformed into different realizations for different purposes, e.g. latent variable profile estimation, process monitoring, fault detection, etc., as discussed in the paper. Copyright © 2003 John Wiley & Sons, Ltd.


  • PLS score–loading correspondence and a bi‐orthogonal factorization

    by Rolf Ergon

    It is established industrial practice to use the correspondence between partial least squares (PLS) scores and loadings or loading weights as a means for process monitoring and control. Deviations from the normal operating point in a score plot are then related to the influences from major process variables as shown in a loading or loading weight plot. These relations are often presented in a bi‐plot, i.e. appropriately scaled scores and loadings or loading weights are displayed in the same plot. As shown in the present paper, however, the orthogonal PLS algorithm of Wold gives no direct theoretical and graphical correspondence, i.e. the bi‐plot will show an angle deviation that causes an interpretational problem. The alternative non‐orthogonal PLS algorithm of Martens gives direct correspondence, but the correlated latent variables may then cause another interpretational problem. As a solution to these problems, this paper presents a PLS factorization where both scores and loadings are orthogonal (BPLS), and we show how the Wold and Martens factorizations can easily be transformed to this solution. The result is independent latent variables as well as direct score and loading correspondence. It is also shown that the transformations involved do not affect the predictor found by PLS regression. The score–loading correspondence properties for the different PLS factorizations are discussed using principal component analysis (PCA) as a reference case. An example using industrial paper plant data is included. Copyright © 2002 John Wiley & Sons, Ltd.


  • Open Access

    Dynamic system multivariate calibration based on multirate sampling data

    by Rolf Ergon, Maths Halstensen

    Least squares regression (PLSR) algorithms based on latent variables (LV) modeling are effective tools for handling ill-conditioned regression data. In many process related cases the data form time series, and it may then be possible to improve the prediction/estimation results by utilizing the autocorrelation in the observations. This can be done by use of estimators found from experimental data by use of a combination of statistical/chemometric and system identification methods. In important industrial cases, the response variables are product qualities which also in the experimental data are sampled at a low and possibly irregular rate, while the regressor variables are sampled at a higher rate. After a discussion of the options available, the paper shows how the autocorrelation of the regressor variables in such multirate sampling cases may be utilized by identification of latent variables based output error (LV + OE) estimators. An example using acoustic power spectrum regressor data is finally presented. 1. Introduction and problem statement In many important practical cases there is a need to estimate primary system responses y i from known system inputs u and secondary system outputs y 2. In typical industrial cases the y i responses are product quality measurements that cannot be made available on-line, while u are manipulated process inputs and y 2 are various on-line process measurements. Dynamic estimators for this purpose may be identified from experimental u, y i and y 2 data (Ergon, 1999b). The experimental data are often characterized by few observations and a large number of nearly collinear y 2 variables. In the static case the statistical/chemometric methods PCR (Principal Component Regression) or PLSR (Partial Least Squares Regression) based on latent variables (LV) models provide powerful tools for handling such ill-conditioned problems. The experimental u and y 2 data are then collected in an X matrix, while the y i data are collected in a y vector (assuming a single response variable), and the X information is compressed into an estimated latent variables matrix T. In many process related cases the X and y data in ordinary PCR or PLSR form time series, and it may then be possible to improve the prediction/estimation results by utilizing the autocorrelation in the observations. After compression of the X data into T, this can basically be done by identification of FIR (Finite Impulse Response), ARMAX (AutoRegressive Moving Average with eXtra inputs), ARX (AutoRegressive with eXtra inputs) or OE (Output Error) type of models. The estimator may thus be found from experimental data by a combination of multivariate calibration methods used in statistics/chemometrics (e.g. Martens and Næs (1989) and Esbensen (2000)) and system identification methods (e.g. Ljung (1999), Söderstrøm and Stoica (1989) and Van Overshee and De Moor (1996)).