Similar Questions

• How to track the progress of Levenberg-Marquardt

  Hi all

  My Levenberg-Marquardt works perfectly, but one iteration takes quite long, about 5 seconds, as expected from the complexity of my problem (~ 50 parameters fit).

  Question: How to display the current number of interaction in the main program, for the user to know that it is still ongoing?

  (My solution so far is to put a probe in the Levenberg-Marquardt subroutine provided on this 'i', which is the current interation. "But it's not easy to use).

  Thanks & best regards!

  Peter, LMU Munich, Germany

  You can use your model VI to increment a counter that appears elsewhere. If your model compute partial derivatives, then there will be only one call per iteration LM model VI plus 1 for the estimate.  If you are not calculating the partial derivatives LM then uses central difference method to calculate them.  This means function by iteration of LM 3 calls, plus 3 for the estimate.

  -Jim

• Levenberg-Marquardt optimization question

  Hello

  I am trying to use the Levenberg-Marquardt algorithm, using a technique of optimization error percentage instead of the technique of the absolute error. AFAIK the Levenberg Marquardt in Labview algorithm uses the technique of the absolute error (i.e. Y - f(x)) when calculating on the errors. How can I change this option to use the technique of percentage error, when the error = [(Y - f(x)) / y] * 100? I changed the calculation of the error in the algorithm to calculate the error in this way, but the results were not satisfactory. I wonder if I forget something or there at - it much more to what meets the eye? I know that in other programs like Mathematica you can choose how you want the error calculated, but I do not see this option in the Labview vi. Any help appreciated.

  Thank you

  LMS

  Additional information:

  I'm using Labview version 8.2

  I changed the Subvi ABX to calculate the error differently. The ABX vi contains a Subvi called gradient and function of LM. The output of this Subvi calculates better suited like Y - the error. I changed this calculation to be [((ajustement Y-meilleur) / y) * 100] to obtain the percentage of error instead. This affects alpha, beta and chisqr. I think I may need to make other changes to the algorithm to implement the technique of error in percentage but I don't know what changes to make. The results that I have received the above changes were worse than the results I got from the method of the absolute error (error = Y – the better suited). This should not be the case. I implement the VI of the Levenberg-Marquardt algorithm format and not the string of the formula. Any help appreciated.

  Thank you

  LMS

  Hi LMS.

  I don't know if I understand.

  However, as the Lev - Mar algorithm calculates Y - f (x). If you divide that result by the best linear fit (Y) and multiply by 100. This will give you your desired effect.

  I hope this helps you.

• Levenberg Marquardt adjustment: problem of formula

  Hello

  I'm so new to LM did I maybe a stupid question... Please bear with me!

  I try to adapt a simple formula but routine LV tells me my formula wrongly.

  is attached a PNG for my case!

  Thanks for any help

  concerning

  N

  Without looking at the help, I'm not sure but check if the formula is case-sensitive.

  Lynn

• Adjustment of curve using Levenberg-Marquardt algorithm

  Hi all

  Thank you for my request answered in the last post. One thing I like about Labview is its ability to allow complex models to be fitted to experimental data with so much ease. I have been using the L - M unconstrained editing for some time. However, there is a question to which I thought I would like to be answered by Labview professional themselves.

  When the curve adjustment is completed, the algorithm always has the minimum value for a reverse parameter. for example. Suppose that a parameter X is found to have a value of 0.8 after editing by l.-m. algorithm. However in principle several orders of values are possible for x. for example is not a sensitive parameter in the main equation so it can have values of 1000 or 400, or 60, 1. In this scenario the algorithm will always give an answer like X = 1; This is the minimum solution?

  See you soon,.

  Navneet

  Navneet,

  Implementation of Lev - Mar in LabVIEW does not explicitly attempt to diminish the importance of any parameter.  If the parameter has as little impact on the distance between the data and the best curve fit, then it will change according to the gradient of the problem.  If the parameter does not affect the distance, so that you can get a linear solver error indicating a deficiency in a row.

  You can detect this behavior by using the Intervals.vi to mount curve nonlinear.  An interval that is very large compared to the other parameters would indicate a low dependence on this setting.

  -Jim

• Adjustment of Levenberg-Marquardt tweaks

  Hello

  I have two questions about the non-linear curve fit LM VI:

  (1) does anyone know why after one iteration successfully the lambda value is multiplied by 0.4 instead of 0.1 as in the implementation of the standard (i.e. Numerical Recipes)? This factor has an influence on the speed of the adjustment?

  (2) is it possible to use analytical derivatives to speed the evaluation of the slope and the Hessian of Chi-square?

  Thank you

  Bernard

  (1) maybe it's so nobody is accusing them of use directly the NR without attribution.  That the value is just an empirical assumption, so there is nothing too magical about 0.1 or 0.4 (better to be less than 1).  As you have found where this value is used (not too hard), I suggest that you experience in the changing of 0.1 or 0.2 or 0.5 and see if that makes a difference.  Let me know, I have not tried.  When I implemented the systematic NR in LV, I used 0.1 and it worked, and when I use the LV routine with 0.4 it still works.

  (2) If you select f(x,a) is a VI a slot - VI is used to calculate the values of service, you can also return the analytical derivatives.  This can save time and reduce rounding error, and I recommend warmly when possible.

• Reference of Levenberg Marquardt

  Hi all

  I implemented the LMA succesfully in my thesis and calculated the necessary uncertanities descrobe by Jim and Altenbach method in an another messages. Jim the guy from the DSP has also shown using NIST standards that the multiplication of the matrix of final covariance with the MSE provides good estimates of the variances. I wanted to refer to it in my thesis and I was wondering if there is any report or document published someone made this list.

  Thank you

  (Kyle, you two last links are broken.)

  Chapter 15 digital revenue has useful information.

• Nonlinear worthy of the equations of the diode

  Hello

  I want to make an adjustment not length of the features of diodes.

  The function that I need to set up is could be written as Y = f (X, Y).

  This means that the data is both the goal of adjustment and a part of the equation.

  Is this possible with the vi made nonlinear of levenberg-Marquardt?

  Best regards

  Baptist

  OK, I understand.

  I ' l try tomorrow morning.

  Best regards

  Baptist

• Call LabVIEW c# DLLs

  Hello

  I need to use the Levenberg-Marquardt algorithm in c#. There is a VI in LabVIEW with Levenberg-Marquardt algorithm and is not able to Studio (2008). I decided to create DLLs in LabVIEW and call in c#.

  So I created LevenbergMarquartFit.vi and Function.vi. I also have Test.vi. In Test.vi LevenbergMarquartFit.vi are called as vi and function. It works correctly under vi and it does not work properly as a function call.

  I also try to call c# fuction.

  Header file LevenbergMarquartFit.h is

  ========

  void __cdecl LevenbergMarquartFit (double tolerance, int32_t MaxIterations,

  Double Y [], double X [], OffsetIn double, double B2In, double A2In,.

  Double B1In, double A1In, int32_t * iterations, double [] did, double * offset,.

  Double * B2, double * A2, double * B1, double * A1, double * residues.

  the mistake of char [], int32_t len2, int32_t len3, int32_t len4, int32_t len);

  ========

  I'm calling from c#

  =========

  Double [] AdYTimeBase = new double [64];

  ...

  Double dTolerance = 0.00000001;

  int iMaxIterations = 1000;

  Double [] AdY;

  Double dOffsetIn = Double.Parse (textBox_Offset.Text) * AdYTimeBase [iTimeBasePoints - 1];

  Double dB2In = Double.Parse (textBox_B2. (Text);

  Double dA2In = Double.Parse (textBox_A2. Text) * AdYTimeBase [iTimeBasePoints - 1];

  Double dB1In = Double.Parse (textBox_B1. (Text);

  Double dA1In = Double.Parse (textBox_A1. Text) * AdYTimeBase [0];

  int iIterations = 0;

  Double [] AdFit = double [new AdYTimeBase.Length];

  Double dOffset = 0;

  Double dB2 = 0;

  double dA2 = 0;

  Double dB1 = 0;

  double dA1 = 0;

  Double dResidue = 0;

  strerror(errno) string = "";

  iLen int = 0;

  int iLen2 = 0;

  int iLen3 = 0;

  int iLen4 = 0;

  LevenbergMarquartFit (dTolerance, iMaxIterations, AdYTimeBase, AdX,

  dOffsetIn, dB2In, dA2In, dB1In, dA1In,

  on iIterations, on AdFit,.

  on dOffset, on dB2, out dA2 on dB1, out dA1.

  on dResidue, out strerror(errno) out iLen, of the iLen2, the iLen3, the iLen4);

  ==========

  but the program hangs at the LevenbergMarquartFit() function.

  Any help?

  Thanks, Andrey.

  Andrey.

  You don't mention what edition of Measurement Studio you use, but if you have publishing company, you should be able to use the CurveFit.NonLinearFit method.  Internally, it uses the Levenberg-Marquardt algorithm.

  NickB

  National Instruments

• Problem of Fit 2D Gaussian.

  I have a 2d Gaussian curve fitting problems.

  I have a set of data: with 3 columns and rows of N. For example

  x y intensity.

  ..   ..      ..

  N is the same for all 3 columns. I'm trying to create a plot intensity out of these data and adapt a 2D Gaussian to it.

  I tried to do a polynomial interpolation of the intensity of data on the x - go and carry out a plot of intensity (do not know if there is a better way to do it! Please Suggest if there is nothing else.).

  Then I use the Levenberg-Marquardt method to adapt to these data. For some reason, it does not any iteration estimation. The best value of adjustment is the same as the estimate. There is shoing error: NI_Gmath.lvlib:LM is coefficients.vi:1 again. (see attached screws)

  Can someone help me with this!

  Thank you.

  OK, I quickly changed my old example and it seems very easily mount your raw data. See attached (LV2010). Modify if needed.

  (Note that I set the width as 'width at half the height', your definition may differ.)

  

• Easier nonlinear never - made feel like I'm going crazy

  Hey everybody,

  I used the non-linear curve of Levenberg-Marquardt mounting screws before with great success.  I've used a few times to adapt to complex functions with many free parameters and functions have always shone until I've created my design vi, limits and conjecture of initial setting well.  Now, I try to adapt the simplest function never and I am bombing badly.

  I have exactly 1 (ONE!) setting I'm trying to adapt.  The function that he spends is dead simple.  But how it works makes me crazy:

  1. no matter what I choose as my initial setting, I still get exactly the same result on the table "better adjust the parameters", even if by choosing different values manually, I can clearly a better fit and the residual variations

  2 no matter what, he always runs exactly 6 iterations.  This happens even if I set the tolerance of the extremely small termination, or even if I set my maximum iterations less 6 Nb!
  3. my best value will be 1e14-1e17 I think.  I was wondering if the reversal in orders of magnitude has been causing crashing the algorithm (maybe because he didn't pass the initial value enough even see a difference for the first couple of iterations or something), so I tried also to redo my function as two parameters: one being the coefficient of a number in scientific notation (with the limits between 1 and 10) and the other being the exponent (with the limits between 1 and 20).  I then treated these two parameters in the model turn them back into a single number and use THIS number in the function, but it still bombed just as badly.  I restored the example I posted to the parameter single original fit.

  I'm just completely baffled.  I admit that it is the first time that I tried to use these screws to adapt to a single free parameter.  These screws are designed only to maintain multiple parameters?  If so, why not a not my experience that I described in (3) above works or at least do something?  Also the turnaround in values causes me trouble?  I feel like I fit stuff with huge fluctuations of values before.

  The fact that I always get exactly 6 iterations, is what really makes me feel the nuts.  How is she even learning to 6 if I set the maximum number of iterations to less than that?

  I searched the forum for the last two hours looking for something like this.  Honestly, I could not find anything.  Sorry if it's a crazy stupid question that has been answered before.

  Help much appreciated!  Thank you!

  "1. no matter what I choose as my initial setting, I still get exactly the same result on the table"better adjust the parameters", even if by choosing manually different values, I can clearly a better fit and the residual changes."
  I tried to manually change the setting and found that the smallest obtained residue if the parameter has become closer to 0. In fact, I saw no change in the residue for the 37 lower parameter values or if, because of the scale of the problem. What is a better fit for you?

  The fact that I always get exactly 6 iterations, is what really makes me feel the nuts.
  Feature County includes the function for digital calculations derived partial calls, and we do not verify the number of function for termination until after the partial derivatives are calculated. By default for digital derived is central difference.

  Partial derivatives are very small in the scale of this problem.  Looks like the stepsize at each iteration is only one or two orders of magnitude above machine epsilon. Due to the relative change very low residue induced by small step size, Lev - Mar may report that it is in convergence. Strongly recommend that the scale of the problem, but this may be a moot point because a manual search for the setting seems to show that the setting tends to 0.

  -Jim

• Adjustment of the non-linear curve - maximum iterations exceeded (error-23026)

  Hello

  In my application, I use the curve non-linear adjustment VI (Levenberg-Marquardt) to adapt the data acquired continuously using a task DAQmx. Often in the course of the implementation of the instrument, the curves are not yet "adaptable" and max iterations set in the parameters of expiration is reached.

  It's OK but the problem is that it generates an error (-23026) that stops the execution of the VI.

  How can I ignore this error and continue to run the VI until the termination is controlled by the setting of tolerance?

  Kind regards

  Bernard

  Hi Bernard,.

  Need to wire the error output and handle the error programmatically.

• Table 1 d coupling to a double-peak Gaussian function

  Hi experts LabVIEW!

  I am rewriting a MATLAB code in LabVIEW code. I'm stuck at this next part:

  ---------------------------------

  (MATLAB code)

  x = 1:770

  y = current_frame_1D

  Options. StartPoint = [amp1, pos1, Lim_mean1 (3), amp2, pos2, Lim_mean1 (3)];

  options = fitoptions ('gauss2', 'Down', Lower_Lim1 (Lower_Lim2) and 'Upper', [Upper_Lim1, Upper_Lim2], 'TolFun', 1e-2, 'TolX', 1-1);

  G_fit = fit(x.',y','gauss2',options);

  ----------------------

  The parameter gauss2 in MATLAB fit function is intended to adapt the data of the following equation:

  Y = a1 * exp (-((x-b1)/c1) ^ 2) + a2 * exp (-((x-b2)/c2) ^ 2)

  That is a sum of two Gaussian curves.

  Is there a way to do this in LabVIEW? (I see that the fit.vi pic Gaussian, which can detect 1 Ridge only...)

  Thank you!

  The Palette of implementation has a non-linear curve adjustment function that uses the Levenberg-Marquardt algorithm to fit an arbitrary (in your case, the sum of two Gaussians) function.  Curve not linear are something of a black art - if you have not used before LM, I'd do a little reading on this topic (use the Web).

  Another method to make such an adjustment is the Simplex method.  There is a downhill Simplex nD VI on the optimization Palette, but when I tried to figure out how to use that to a problem of optimization/smoothing of mine, I gave up and wrote my own the simplex algorithm in LabVIEW.  I know not if it would work for your data and model - if you can attach a file of data, I could give it a try...

  Bob Schor

• Estimate of exponential decay time constant - taking tangent lines

  Hello

  How would you estimate a first order decay time constant? I have a system that loses heat very slowly and I wish assess the time constant of the first part of the data in order to allow more frequent tests (I know it will end by decomposing to).

  I know I can make a good estimate based on the initial slope of the curve, but I don't know how to implement this in LabVIEW. Does anyone know how I can make the VI calculate a tangent line to a graph, and to find where it crosses another line?

   

  
  Thank you

  How about using the

  VI exponential form

  Have the Palette: Screw

  Request: Full development system

  Returns the exponential fit to a set of data (X, Y) using the method of least squares, less absolute residual or Bisquare.

  This VI uses the General iterative least squares method and the method of Levenberg-Marquardt to adjust the data to a curve exponential of the General form described by the following equation:

  f = ae^bx + c

  where x is the input sequence X, a is the amplitude, b is depreciationand c is out of step. This VI find the values of a, b, and c that better observations (X, Y).

• non-linear regression

  Hello

  I need to make a non-linear regression in LabWindows but I don't not eat in order to find a function for that...

  I know that there s for linear regression (LinFit) function, but not could find one for non-linear... Is there such function CVI?

  Or maybe someone has an algorithm to do this...

  Any help is very appreciated!

  Thank you

  ZZippo

  Hi ZZippo,

  as long as you have the AdvancedAnalysisLibrary there are three related functions for nonlinear curve fit:

  NonLinearFit (-);

  NonLinearFitWithMaxIters (, 100);

  NonLinearFitWithWeight (, 100);

  the last two offer more options than the first.

  If you do not have this library (i.e. the full version of the CVI instead of the basic version), you can use the following implementation of Levenberg-Marquardt:http://www.ics.forth.gr/~lourakis/levmar/