wilmott.com

It would save me typing it in TIA.

I think that the Reference for Levenberg-Marquardt is in "netlib" www.netlib.org in the packet "minpack" but it is in fortran...

audetto, thank you for answering. I made up my mind to translate a Maple solution: it would cover my requirements and would give me a base for intrinsic testing.

Find enclosed a solution for Excel. It is a little bit dirtyand small, but it works. In the enclosed sheet i restrict toonly a few (16) observations (since dynamic arrays are awfullin Excel, but can be done), ignored weights (due to speed) andused a logistic function as test.Due to L-M one has to choose a rough but not completely absurdinitial guess for the parameters.Excel can not compute symbolic or reasonable parse expressionsi use 'precomputed' functions and gradients (which is ok forme). For a different parametric function one has to adapt fand gradf within the VBA source and possibly the number ofparameters (and observations).I also added a protocol from the original Maple solution (sincethe easiest is to let an CAS do the work and link it to Excelgetting the desired flexibility) showing the statistics for it:it is an example that a good statistics is not good enough.

@avtis there any chance that you could walk me through your excel macro?

äääh ... it is the VBA translate of the Maple code (statistics kicked out, dim=1, static length)... i will try to understand it again ... send me a private mail, address is in my profileadded: David Holmgren once gave me that link to look at the section "Solving for viewpoint and model parameters"

Avt,I was trying to use your code and changed the initial p1 parameter, but the VB gave me an error "compilation error - cannot assign to a table" for tmp_beta parameter.any thoughts?

ymikhale,What exactly did you change where? Changing the input in the shiet i can not reproduce it (except forcing the algo to fail).The error message typically occurs, if you change the source code. For this please note that p_start is thought as a vectorin IR^3 (for the parameters of f) and f(xi, tau1, tau2, tau3) is a 1-dim fct. So take care to change everything consistently.tmp_beta is the matrix gradF*gradF^(transpose) [and depends in size on f and data feed in]. Perhaps you may checkwhether matrix operations work within your Excel in a sheet first and make sure that 'all dimensions fit together' usingpaper and pencil

Avt, the only thing I changed was the p_start parameter p1 from 2000 to 1000 in the spreadsheet. At first I thought it was the MMult function that it did not recognize, but then it worked fine for beta 2 lines before. So I am at a loss. Is it possible it has something to do with me having the French version of Excel ?

Hmm ... tmp_beta is not what i wrote, it is a vector, so you may work around looping through it explicitely.But for me entering 1000 works. It should not depend on language versions (even if i have a german one).I dont have a good idea and no TechNet CD at hand. May you wish to store after entering '1000' andupload the thing? I will look at it in the evening.

Hi Avt,I have attached 2 files - one is the original code that I downloaded from the site (***test2), the other is the one where I tried to work around tmp_beta and alpha (it seems that it has a problem with variables declared as variants with ()). None of them worked.Really appreciate your help since I am not a VB expert...

Avt,Here are the files

ymikhale,The shiet works, on different PCs with different operating systems and different Excels ...The last idea i have on it: may be you miss references to libraries.Within VBA check through the menu bar tools / references whether you havethe following 4 'standard ones' activated: Visual Basic for Applications,Microsoft Excel 9.0 Object Library (EXCEL9.OLB), OLE Automation (not need) and Microsoft Office 9.0 Object Library (MSO9.DLL).If it works: fine. If not ... then send me a private mail, may be i look for work arounds