New Version 6.1.3 – New Calibration Method

The new version includes a new option for calibration, based on the traditional Nelder Mead optimisation method. This method is mainly used to obtain the best parameter values based on an objective function. This method is based on another philosophy compared to the conventional Monte Carlo Method (GLUE) or the Bayesian Calibration. Instead of searching for a range of acceptable candidates the Nelder Mead approach will try to find the best possible candidates.

The selection of Nelder Mead is part of the MultiRun methods and can be find on the MultiRun options and used for any of the included parameters.

Menu with use of Nelder Mead

The new method is using similar like the Bayesian Method a range and a start value for each of the parameters. Note that the Range is not strictly inside this range but those values are used to start the search for the best parameter candidates. The objective function can be selected by using the sum of square error or optionally an information criteria index that is also accounting for the number of data points and the number of parameters.

The Index have been suggested within a calibration project currently ongoin within the AgMip organisation to improve the use of agricultural crop models.

Note that the objective function is making a summation of all validation variables and this summation is now added as an output together with the calibrated parameters. See below:

Output results table from Completed MultiRun using Nelder Mead method
Showing how the finding of the parameters values are converting together with two information criteria ouputs. Note that the Run Numbers in this example is multiples of 12 since the SSE is a summation from 12 runs covering a serie of sites. So the actual chain for searching of candidates is here 200. Normally the Nelder Mead requires something like 200 steps to find an optimal solution.

Leave a Reply

Your email address will not be published. Required fields are marked *