Forgot Password | New Account

Close This Window

Merci de télécharger les logos officiels de l'ILL ici


Pour utiliser sur le web

Pour imprimer en haute résolution


Télécharger PNG

Télécharger TIF


Télécharger JPG

Télécharger EPS



Computing for Science

The Computing for Science (CS) group supports ILL scientists, students and visitors in a number of activities including data analysis, instrument simulation and sample simulation.

Back to ILL Homepage
English French Deutsch 

All Software

Version 4.2. (1999)

Basic fitting

By now you should have successfully loaded a set of data, and chosen a fitting function. There should be three Mfit windows open: the control window, the data window, and the parameters window. You can now begin fitting.

1. Get starting parameters

Non-linear fitting works by using an algorithm to attempt to improve on a first 'guess' that you make for the function parameters that fit your data. The first step in fitting is to give Mfit a set of starting parameters to work with. You can do this in two ways. The first is simply to type them into the parameters window:

Fig 1 : Mfit : parameter window

you click on the box showing the value of the parameter you want to change, and type the new value. After you have entered values for all the parameters, press the 'plot' button in the control window, and the function will be drawn in the data window.

This is slow, and for many functions it's not obvious what starting parameters might be. Many Mfit fit functions allow you to get the starting parameters with the mouse, by clicking on points in the data window. Click on the 'guess' button in the control window:

Fig 2 : Mfit : control window asking for a peak position.

and follow the prompts displayed in the message box, the wide test box just above the row of buttons in the control window. For example, many peak fit functions ask you to click on the centre of the peak (in the data window), then on the width (i.e. the half height position), and then on the background. When you've fiinished, the function will be draw in the data window.

2. Beginning a fit

Having chosen starting parameters, you can now begin the fit by pressing the 'fit' button in the control window. As the fit progresses Mfit shows the number of fit iterations in the message box, and draw the function with the current parameters in red in the data window. When the fit finishes, the word 'Done' is displayed in the message box, and the values in the parameters window are the results of the fit. The parameters window also displays the uncertainties associated with each parameter value, and the value of the reduced chi squared, and Q, the probability that chi squared should exceed its current value by chance.

Next: Saving results

More advanced fitting

1. Fix and free parameters

You will often want to fix some of the function parameters to preset values, preventing the fitting algorithm from changing them during the fit. You do this simply by clicking in the check box to the left of the parameter name in the parameters window:

Fig 3 : Mfit : parameter window with a fixed backgroung

Here, the 'Background' parameter has been fixed at 9.06. You can free a parameter by clicking on the checkbox again.

2. Select data to fit

Sometimes, you may want to restrict the data points you want to fit: you might just want to fit a small feature in your data, or you may want to exclude data points that you believe are spurious (dangerous, unless you have very good reasons!) You can include and exclude points from the fit by choosing operations from the 'select data' menu in the data window:

Fig 4 : Mfit : Deselected points in the Data window

To select data points to include in the fit, choose 'Select Data' in the 'Data' menu of the 'Mfit : Data' window . The cursor will change to a cross hair, and you click and drag on the data window to define a box enclosing the data points you want to select. Keep clicking and dragging until you have selected all the points you want to include. You deselect points in the same way. To set the zoom mode again (default mode), just use the Zoom menu.

3. The AutoGuess feature

Most fit functions require peak positions, widths, amplitudes and backgrounds. Mfit can perform an automatic guess for such parameters. When checking the 'AutoGuess' item in the Option menu of the control window, it analyzes the data, looking for peak characteristics, and sets the guessed values to parameters, according to their names. It does work fine for most functions, but the data must contain enough points to enable an efficient data analysis.

4. Change fit routine options

It is possible to change some parameters controling the fitting process. To do so, select the Fit Control item of the Options menu in 'Mfit : Control window':

Fig 5 : Mfit : The Fit control window

You can there change :

  • dp - this parameter depends on the fit routine  (see Choosing an alternate fit routine)
    • for the Marquadt Levenberg method, this is the step used to compute partial derivatives <tt>f(x,p+dp) - f(x,p)/dp</tt> (in percent of parameter). Reasonable value should be below 0.1 (10 %) ; when higher, this could lead to uncertainties in derivative estimates.
    • for the Simplex method, it indicates the exploration range around starting parameter set (in %). For instance, a value of 0.1 will indicate that each parameter will first be investigated between +10%/-10% around starting value. This range can be extended during simplex fitting process. Reasonable values should be between, say 0.1 (10 %) and 0.5(50 %). High values are to be used when your starting parameters are very approximative.
  • niter - maximum number of iterations
  • stol - sets the convergence tolerancy for fit. When the fit criterion (usually least-square) variation between two fit iterations is less than that value, the fit is considered to be stable, and finishes.

5. Choosing an alternate fit routine

The standard fit procedure (optimization) is to minimize the least-square criterion by mean of an adaptative gradient (Marquardt-Levenberg method). This routine is supplied in a graphic mode (showing the fit process, and the converging parameters), as well as a non-graphic faster routine.
But you may choose to use some other fitting procedure in the 'Fit Routine' menu of the Control window. For instance, a simplex method is also supplied.

  • Marquardt-Levenberg method - is fast and efficient, but can converge towards a wrong solution in some cases (local criterion minimum)
  • Simplex method - is slower, but can easely avoid local criterion minima solutions, and finally converge to the global minimum

Refer to the How to write fitting routines page in order to add your own minimization method.

Next: Saving results
See also: How to write fitting routines , How to write fitting functions