From 63e0ce4866e63cc6a844bfcb7b98c98619dd2831 Mon Sep 17 00:00:00 2001 From: Johannes Ranke Date: Tue, 14 Oct 2014 20:36:36 +0200 Subject: Add a note about transforming fractions --- vignettes/gmkin_manual.Rmd | 8 +++++++- vignettes/gmkin_manual.html | 3 ++- 2 files changed, 9 insertions(+), 2 deletions(-) (limited to 'vignettes') diff --git a/vignettes/gmkin_manual.Rmd b/vignettes/gmkin_manual.Rmd index 8dd4c54..39669e8 100644 --- a/vignettes/gmkin_manual.Rmd +++ b/vignettes/gmkin_manual.Rmd @@ -282,7 +282,8 @@ plot of the model and the data using the specified initial parameter values. If a similar model with a partially overlapping model definition has already be fitted, initial values for parameters with the same name in both models can also be retrieved -from previous fits. This facilitates stepwise fitting of more complex degradation pathways. +from previous fits by selecting the fit and pressing the button "Get initials +from". This facilitates stepwise fitting of more complex degradation pathways. After the model has been successfully fitted by pressing the "Run" button, the optimised parameter values are added to the parameter table. @@ -310,6 +311,11 @@ as defined in the model, or if they are internally transformed during the fittin order to improve the estimation of standard errors and confidence intervals which are based on a linear approximation at the optimum found by the fitting algorithm. +If fitting with transformed fractions leads to a suboptimal fit, doing a first +run without transforming fractions may help. A final run using the optimised +parameters from the previous run as starting values (see comment on "Get +initials from" above) can then be performed with transformed fractions. + The dropdown box "weight" specifies if and how the observed values should be weighted in the fitting process. If "manual" is chosen, the values in the "err" column of the dataset are used, which are set to unity by default. Setting these to higher values diff --git a/vignettes/gmkin_manual.html b/vignettes/gmkin_manual.html index 63ed263..cd66327 100644 --- a/vignettes/gmkin_manual.html +++ b/vignettes/gmkin_manual.html @@ -182,7 +182,7 @@ if (window.hljs && document.readyState && document.readyState === "complete") {

Parameters

In the right hand area, initially the tab with the parameter list is displayed. While name and type of the parameters should not be edited, their initial values can be edited by clicking on a row. Also, it can be specified if the parameters should be fixed in the optimisation process.

If the initial values for the parameters were changed, the resulting model solution can be visually checked by pressing the button “Show initial”. This will update the plot of the model and the data using the specified initial parameter values.

-

If a similar model with a partially overlapping model definition has already be fitted, initial values for parameters with the same name in both models can also be retrieved from previous fits. This facilitates stepwise fitting of more complex degradation pathways.

+

If a similar model with a partially overlapping model definition has already be fitted, initial values for parameters with the same name in both models can also be retrieved from previous fits by selecting the fit and pressing the button “Get initials from”. This facilitates stepwise fitting of more complex degradation pathways.

After the model has been successfully fitted by pressing the “Run” button, the optimised parameter values are added to the parameter table.

@@ -192,6 +192,7 @@ if (window.hljs && document.readyState && document.readyState === "complete") {

The “solution_type” can either be “auto”, which means that the most effective solution method is chosen for the model, in the order of “analytical” (for parent only degradation data), “eigen” (for differential equation models with only linear terms, i.e. without FOMC, DFOP or HS submodels) or “deSolve”, which can handle all model definitions generated by the mkin package.

The parameters “atol” and “rtol” are only effective if the solution type is “deSolve”. They control the precision of the iterative numerical solution of the differential equation model.

The checkboxes “transform_rates” and “transform_fractions” control if the parameters are fitted as defined in the model, or if they are internally transformed during the fitting process in order to improve the estimation of standard errors and confidence intervals which are based on a linear approximation at the optimum found by the fitting algorithm.

+

If fitting with transformed fractions leads to a suboptimal fit, doing a first run without transforming fractions may help. A final run using the optimised parameters from the previous run as starting values (see comment on “Get initials from” above) can then be performed with transformed fractions.

The dropdown box “weight” specifies if and how the observed values should be weighted in the fitting process. If “manual” is chosen, the values in the “err” column of the dataset are used, which are set to unity by default. Setting these to higher values gives lower weight and vice versa. If “none” is chosen, observed values are not weighted. Please refer to the documentation of the modFit function from the FME package for the meaning of options “std” and “mean”.

The options “reweight.method”, “reweight.tol” and “reweight.max.iter” enable the use of iteratively reweighted least squares fitting, if the reweighting method is set to “obs”. Please refer to the mkinfit documentation for more details.

The drop down box “method.modFit” makes it possible to choose between the optimisation algorithms “Marq” (the default Levenberg-Marquardt implementation from the R package minpack.lm), “Port” (an alternative, also local optimisation algorithm) and “SANN” (the simulated annealing method - robust but inefficient and without a convergence criterion).

-- cgit v1.2.1