6
$\begingroup$

Final Edit: I have found a solution! Specifying Method->"LevenbergMarquardt", Gradient->"FiniteDifference" gives the behavior I am looking for. Huzzah! It appears that the crashing behavior was coming from some inability of the fitting routines to correctly calculate the gradient. And further, this solution works perfectly without having to correct for the differing magnitudes of the parameters or accounting in any explicit way for the overparamaterization.

Another Edit: Added full context of the problem to the end of the post.

Edited to add: The choices of Method given in the comments and in the answer provided by Tim work for the example I originally provided, and I thank all of you for those suggestions. Unfortunately, trying to expand that solution to my more-complex problems also fails. I am therefore adding a more complete example of the problem (the original question will be retained at the end).

First, some sample data to fit to:

sample[t_] = (0.002 + 101 t - 461000 t^2 + 2.218 10^9 t^3 - 
             3.64 10^12 t^4 + 3.17 10^15 t^5) Exp[-8653 t];
data = Table[{t, sample[t] + RandomVariate[NormalDistribution[0, 0.00001]]},
             {t, 0, 0.002, 0.000004}];
ListPlot[data]

enter image description here

Now the model:

rateeqs = {a'[t] == k1b b[t] + ksqb b[t] a[t] + kttb b[t]^2 + 
                    kbd b[t] c[t] - kdb a[t] d[t] , 
           b'[t] == -k1b b[t] - ksqb b[t] a[t] - kttb b[t]^2 - 
                    kbd b[t] c[t] + kdb a[t] d[t] , 
           c'[t] == k1d d[t] + ksqd d[t] c[t] + kttd d[t]^2 + 
                    kdb a[t] d[t] - kbd b[t] c[t], 
           d'[t] == -k1d d[t] - ksqd d[t] c[t] - kttd d[t]^2 - 
                    kdb a[t] d[t] + kbd b[t] c[t]};
initconc = {a[0] == a0, b[0] == b0, c[0] == c0, d[0] == d0};
additionaltdeps = {abs60[t] == 5 eps60 b[t], abs70[t] == 5 eps70 d[t],
                   abs[t] == abs60[t] + abs70[t]};
additionalinitcond = {abs60[0] == 5 eps60 b[0], abs70[0] == 5 eps70 d[0], 
                      abs[0] == abs60[0] + abs70[0]};
tdepvars = {a, b, c, d, abs60, abs70, abs};

Setting up the fixed parameters, the variable parameters, and initial guesses for those variable parameters:

fixedparams = {k1b -> 6000, k1d -> 100, ksqb -> 10^6, ksqd -> 10^6, 
               kttb -> 10^9, kttd -> 10^9, a0 -> 4 10^-5, c0 -> 2 10^-5, 
               eps60 -> 3500, eps70 -> 12000};
varparams = {kbd, kdb, b0, d0};
initguesses = {kbd -> 5 10^8, kdb -> 10^8, b0 -> 10^-7, d0 -> 10^-8};

Finding the paramaterized solution:

solution = ParametricNDSolve[Join[rateeqs, initconc, additionaltdeps, 
                                  additionalinitcond] /. fixedparams, 
                             tdepvars, {t, 0, 0.002}, varparams];

Demonstrating that it works:

Show[ListPlot[data, PlotRange -> Full], 
     Plot[((abs /. solution) @@ Values[initguesses])[t], {t, 0, 0.002}, 
          PlotRange -> Full, PlotStyle -> Red], PlotRange -> Full]

enter image description here

And now trying to do the fit:

fitfn = abs /. solution;
tmp = Values@initguesses;
Dynamic@Column[{Show[ListPlot[data, PlotRange -> Full], 
                     Plot[(fitfn @@ tmp)[t], {t, 0, 0.002}, 
                          PlotRange -> Full, PlotStyle -> Red], 
                     PlotRange -> Full, ImageSize -> Large], 
                ListPlot[{#1, #2 - (fitfn @@ tmp)[#1]} & @@@ data, 
                         PlotRange -> Full, AspectRatio -> 0.2, 
                         ImageSize -> Large]}]
NonlinearModelFit[data, (fitfn @@ varparams)[t], 
                  Evaluate[List @@@ initguesses], t, 
                  Method -> "NMinimize", 
                  StepMonitor :> (tmp = varparams)]

I have tried NMinimize, as listed above. In those cases, I get error messages that don't make sense (as one example of many, input value outside the range of data in the interpolating function, for example, when in fact the given data point is right smack in the middle of the interpolating function's range). And my Dynamic display of the fitting progress never updates.

I have tried various NDSolve formulations, in which case the kernel seems to quit with no error message.

Original Question Below

I am finding some inconsistencies in getting NonlinearModelFit to work with the output of ParametricNDSolve. Here is an example that works (starting with a fresh Kernel):

eqs = {a'[t] == -k1 a[t] - k2 a[t]^2, 
       b'[t] == k1 a[t] + k2 a[t]^2, 
       a[0] == a0, b[0] == b0};
fixedparams = {k1 -> 1.2, b0 -> 0};
fns = {a, b};
params = {k2, a0};
solution = ParametricNDSolve[eqs /. fixedparams, fns, {t, 0, 5}, params]
fitfn = a /. solution;
paramsForDataSet = {k2 -> 1.263, a0 -> 0.0321};
dataset = {#, ((fitfn @@ params) /. paramsForDataSet)[#] + 
  RandomVariate[NormalDistribution[0, 0.0002]]} & /@ Range[0, 5, 0.01];
ListPlot[dataset, PlotRange -> Full]

enter image description here

initialGuess = {k2 -> 2.0, a0 -> 0.3};
tmp = Values@initialGuess;
Dynamic@Column[{Show[ListPlot[dataset, PlotRange -> Full], 
                     Plot[(fitfn @@ tmp)[t], {t, 0, 5}, 
                          PlotRange -> Full, PlotStyle -> Red], 
                  PlotRange -> Full, ImageSize -> Large], 
                ListPlot[{#1, #2 - (fitfn @@ tmp)[#1]} & @@@ dataset, 
                         PlotRange -> Full, AspectRatio -> 0.2, 
                         ImageSize -> Large]}]

This last bit gives me a dynamically-updating plot of my fit and the residuals as it converges. Here is the fitting procedure:

result = NonlinearModelFit[dataset, (fitfn @@ params)[t], 
                       Evaluate[List @@@ initialGuess], t, 
                       StepMonitor :> (tmp = params)]
tmp = Values@result["BestFitParameters"]

enter image description here

enter image description here

This looks great! But when I slightly complicate the model, it crashes the kernel on me. Again starting from a fresh kernel:

eqs = {a'[t] == -k1 a[t] - k2 a[t]^2, b'[t] == k1 a[t] + k2 a[t]^2, 
       c[t] == q a[t] + r b[t], c[0] == q a0 + r b0, a[0] == a0, 
       b[0] == b0};
fixedparams = {k1 -> 1.2, b0 -> 0};
fns = {a, b, c};
params = {k2, a0, q, r};
solution = ParametricNDSolve[eqs /. fixedparams, fns, {t, 0, 5}, params]
fitfn = c /. solution;
paramsForDataSet = {k2 -> 1.263, a0 -> 0.0321, q -> 0.341, 
                    r -> 0.8431};
dataset = {#, ((fitfn @@ params) /. paramsForDataSet)[#] + 
       RandomVariate[NormalDistribution[0, 0.0002]]} & /@ Range[0, 5, 0.01];
ListPlot[dataset, PlotRange -> Full]

enter image description here

initialGuess = {k2 -> 2.0, a0 -> 0.3, q -> 0.32, r -> 0.88};
tmp = Values@initialGuess;
Dynamic@Column[{Show[ListPlot[dataset, PlotRange -> Full], 
                     Plot[(fitfn @@ tmp)[t], {t, 0, 5}, PlotRange -> Full, 
                     PlotStyle -> Red], 
                  PlotRange -> Full, ImageSize -> Large], 
                ListPlot[{#1, #2 - (fitfn @@ tmp)[#1]} & @@@ dataset, 
                  PlotRange -> Full, AspectRatio -> 0.2, 
                  ImageSize -> Large]}]
result = NonlinearModelFit[dataset, (fitfn @@ params)[t], 
           Evaluate[List @@@ initialGuess], t, 
           StepMonitor :> (tmp = params)]
tmp = Values@result["BestFitParameters"]

The only differences are:

  • adding c[t] and c[0] to eqs
  • adding c to fns
  • adding q and r to params
  • adding values for q and r to paramsForDataSet and to initialGuess
  • changing fitfn to c instead of a

Everything else is identical, but this time the kernel crashes. Any suggestions would be most welcome.

(In case this is a bug in Mathematica, I have submitted a bug report to Wolfram. I don't want to rule out, though, that I may be doing something wrong, which is why I am asking here as well.)

Fuller context: The sense I am getting from some of the answers and comments is that the particular problem I am posing is poorly-formed because of the overparameterization. Hopefully this explanation will help explain exactly why I need it to smoothly handle such overparameterization.

I am developing an extension to my Chemistry Package (info available here: http://kevinausman.net). In this extension, I am providing an interactive interface to allow the user to fit experimental data with arbitrarily-complex chemical kinetics mechanisms. Here are some images from the interface:

After entering the chemical kinetics mechanism in standard chemistry notation, the user can select which steps of the mechanism are active:

enter image description here

The package then automatically determines the differential rate equations that result from this sub-mechanism (shown at the top of the screenshot below), and then can add further time-dependent variables (in this case, absorbance 60, absorbance 70, and total absorbance), time-independent parameters (in this case, the extinction coefficients), and equations relating them to the automatically-determined differential rate laws.

enter image description here

The user can then specify a Mathematica symbol that contains the experimental data to be fitting to.

enter image description here

The user then has an interface that allows them to adjust parameters, look at the comparison of the simulation to the data (including residuals), look at a sensitivity analysis of any of the parameters, and then, hopefully, tell the system to go off an try to optimize a set of parameters while holding others constant.

enter image description here

Some of these fixed parameters will be fixed because they are determined by way of separate experiments. Some of them will be fixed temporarily in order to allow one or more of the other parameters to migrate toward a better starting point, and will the later be allowed to vary.

This type of fitting procedure is extremely common in the natural sciences, and is a staple in such scientific software packages as OriginLab, SigmaPlot, and many others. Because Mathematica is much more commonly-available as site licenses at universities, I am trying to develop routines to do the same thing in Mathematica, in a manner that does not require the user to be particularly fluent in Mathematica. So a crashing of the kernel because a particular choice of kinetic model and varying parameters is overparameterized? Not acceptable in this context. It needs to do what it can, recognize when it cannot, and smoothly allow continued operation of the dynamic interface.

I hope that helps explain the context of what I am doing.

$\endgroup$
8
  • 1
    $\begingroup$ Add the option Method->”NMinimize” to NonlinearModelFit and use 0.03 for the initial value of a0. That is more of a work-around than a fix. $\endgroup$
    – JimB
    Aug 17, 2019 at 5:46
  • 2
    $\begingroup$ The use of Method->"NMinimize" was suggested by @UlrichNeumann in more than one answer/comment. That option appears to be more numerically stable than the default method. $\endgroup$
    – JimB
    Aug 17, 2019 at 6:02
  • $\begingroup$ That worked for the system I originally posted, but even that isn't working once I move past my toy system, unfortunately. I will update the OP to include a more detailed example. $\endgroup$ Aug 17, 2019 at 22:36
  • $\begingroup$ I think with your more complex model there are 2 issues: (1) The parameters differ by large orders of magnitude and (2) the model is overparameterized. Change instances of kbd to kbd * 10^8, kdb to kdb * 10^8, etc., along with the necessary changes in the initial values. The model will then converge but the estimators for b0 and d0 seem to be perfectly negatively correlated which is an indication of an overparameterized model. $\endgroup$
    – JimB
    Aug 18, 2019 at 7:02
  • 1
    $\begingroup$ Understood. While we can disagree that the model is over-parameterized for the data generation process, that should not make the kernel crash. Specifically, after a "successful" convergence executing nlm["ParameterTable"] crashes the kernel. I would write WRI with a minimal working example. $\endgroup$
    – JimB
    Aug 22, 2019 at 14:42

2 Answers 2

1
$\begingroup$

Update Right below I give evidence that the model is over-parameterized for the data generation process.

I've put the calculations in a loop and perform just 10 simulations. (1,000 simulations is better but not completely necessary.) One can see that the estimator for kbd is nearly perfectly related to the estimator for kdb. Therefore those two parameters are nearly redundant. That also hinders the underlying algorithm from finding the appropriate estimators.

sample[t_] = (0.002 + 101 t - 461000 t^2 + 2.218 10^9 t^3 - 
     3.64 10^12 t^4 + 3.17 10^15 t^5) Exp[-8653 t];

rateeqs = {a'[t] == k1b b[t] + ksqb b[t] a[t] + kttb b[t]^2 + kbd 10^8 b[t] c[t] - 
     kdb 10^8 a[t] d[t], 
   b'[t] == -k1b b[t] - ksqb b[t] a[t] - kttb b[t]^2 - kbd 10^8 b[t] c[t] + kdb 10^8 a[t] d[t], 
   c'[t] == k1d d[t] + ksqd d[t] c[t] + kttd d[t]^2 + kbd 10^8 a[t] d[t] - 
     kdb 10^8 b[t] c[t], 
   d'[t] == -k1d d[t] - ksqd d[t] c[t] - kttd d[t]^2 - kbd 10^8 a[t] d[t] + kdb 10^8 b[t] c[t]};
initconc = {a[0] == a0, b[0] == b0 10^-7, c[0] == c0, 
   d[0] == d0 10^-8};
additionaltdeps = {abs60[t] == 5 eps60 b[t], abs70[t] == 5 eps70 d[t], abs[t] == abs60[t] + abs70[t]};
additionalinitcond = {abs60[0] == 5 eps60 b[0], 
   abs70[0] == 5 eps70 d[0], abs[0] == abs60[0] + abs70[0]};
tdepvars = {a, b, c, d, abs60, abs70, abs};

fixedparams = {k1b -> 6000, k1d -> 100, ksqb -> 10^6, ksqd -> 10^6, 
   kttb -> 10^9, kttd -> 10^9, a0 -> 4 10^-5, c0 -> 2 10^-5, 
   eps60 -> 3500, eps70 -> 12000};
varparams = {kbd, kdb, b0, d0};
initguesses = {kbd -> 5, kdb -> 5, b0 -> 2, d0 -> -3};

solution = ParametricNDSolve[
   Join[rateeqs, initconc, additionaltdeps, additionalinitcond] /. 
    fixedparams, tdepvars, {t, 0, 0.002}, varparams];

fitfn = abs /. solution;
tmp = Values@initguesses;
SeedRandom[12345];
nSimulations = 10;
mle = ConstantArray[{0, 0, 0, 0}, nSimulations];
Do[data = 
  Table[{t, 
    sample[t] + RandomVariate[NormalDistribution[0, 0.00001]]}, {t, 0,
     0.002, 0.000004}];
 Quiet[nlm = 
   NonlinearModelFit[data, (fitfn @@ varparams)[t], 
    Evaluate[List @@@ initguesses], t, Method -> "NMinimize"]];
 mle[[i]] = {kbd, kdb, b0, d0} /. nlm["BestFitParameters"],
 {i, nSimulations}]

Now plot the estimators for kbd vs kdb for the 10 simulations:

Plot of kbd vs kdb

We see that if one knows kbd, then one knows kdb (at least with respect to the way that the data is generated). It takes two to tango: the model and data generation process go together. For this data generation process kbd and kdb are redundant parameters. Even each simulation has an estimated correlation coefficient of nearly 1.0 for these two parameters.

One of the other consequences of this parameter redundancy is that the estimates of the standard errors from NonlinearModelFit are way too small. For example, the estimated standard error of kbd for each simulation tends to be smaller than 0.03. However the standard deviation of just the 10 estimates of kbd is around 0.8.

The good news is that the predictions aren't particularly affected. An over-parameterized model generally predicts just as well as the appropriately parameterized model. It's just the estimates of the parameters (and the associated standard errors) that one needs to be wary about.

Original response

Note: This response only addresses your updated/more complex model. @TimLaska gave the complete answer to your original question.

I think with your more complex model there are 2 issues: (1) The parameters differ by large orders of magnitude and (2) the model is overparameterized.

Change instances of kbd to kbd * 10^8, kdb to kdb * 10^8, etc., along with the necessary changes in the initial values:

sample[t_] = (0.002 + 101 t - 461000 t^2 + 2.218 10^9 t^3 - 3.64 10^12 t^4 + 3.17 10^15 t^5) Exp[-8653 t];
SeedRandom[12345];
data = Table[{t, sample[t] + RandomVariate[NormalDistribution[0, 0.00001]]}, {t, 0, 0.002, 0.000004}];

rateeqs = {a'[t] ==  k1b b[t] + ksqb b[t] a[t] + kttb b[t]^2 + kbd 10^8 b[t] c[t] - kdb 10^8 a[t] d[t], 
   b'[t] == -k1b b[t] - ksqb b[t] a[t] - kttb b[t]^2 - kbd 10^8 b[t] c[t] + kdb 10^8  a[t] d[t], 
   c'[t] == k1d d[t] + ksqd d[t] c[t] + kttd d[t]^2 + kbd 10^8 a[t] d[t] - kdb 10^8  b[t] c[t], 
   d'[t] == -k1d d[t] - ksqd d[t] c[t] - kttd d[t]^2 - kbd 10^8 a[t] d[t] + kdb 10^8  b[t] c[t]};
initconc = {a[0] == a0, b[0] == b0 10^-7, c[0] == c0, d[0] == d0 10^-8};
additionaltdeps = {abs60[t] == 5 eps60 b[t], abs70[t] == 5 eps70 d[t], abs[t] == abs60[t] + abs70[t]};
additionalinitcond = {abs60[0] == 5 eps60 b[0], abs70[0] == 5 eps70 d[0], abs[0] == abs60[0] + abs70[0]};
tdepvars = {a, b, c, d, abs60, abs70, abs};

fixedparams = {k1b -> 6000, k1d -> 100, ksqb -> 10^6, ksqd -> 10^6, 
   kttb -> 10^9, kttd -> 10^9, a0 -> 4 10^-5, c0 -> 2 10^-5, 
   eps60 -> 3500, eps70 -> 12000};
varparams = {kbd, kdb, b0, d0};
initguesses = {kbd -> 5, kdb -> 1, b0 -> 1, d0 -> 1};
(* initguesses={kbd\[Rule]5 10^8,kdb\[Rule]10^8,b0\[Rule]10^-7,d0\[Rule]10^-8}; *)
solution = ParametricNDSolve[Join[rateeqs, initconc, additionaltdeps, additionalinitcond] /. 
    fixedparams, tdepvars, {t, 0, 0.002}, varparams];

fitfn = abs /. solution;
tmp = Values@initguesses;
nlm = NonlinearModelFit[data, (fitfn @@ varparams)[t], 
   Evaluate[List @@@ initguesses], t, Method -> "NMinimize"];
nlm["ParameterTable"]

ParameterTable

(cov = nlm["CorrelationMatrix"]) // TableForm

Parameter correlation matrix

Show[ListPlot[data], Plot[nlm[t], {t, 0, 0.002}, PlotStyle -> Red]]

Data and fit

The model converges (with some warning messages) and has the appearance of a producing a good fit but the estimators are all highly correlated with each other. That suggests that the model could be overparameterized for the available data. In short, the fit is good but the parameter estimates should not be taken too seriously.

$\endgroup$
8
  • $\begingroup$ 1) Thank you! 2) Some of the warning messages seem bizarre. For example, trying to evaluate the function at points within the defined range, but saying it is outside the range? Odd. 3) Changing the Method to anything other than NMinimize seems to repeat the crashing behavior. I am fine with using NMinimize... it just seems odd that it would fail so badly. $\endgroup$ Aug 19, 2019 at 17:55
  • $\begingroup$ 4) Many scientific problems involve calculations using vastly different orders of magnitude, so it seems strange to me that NonlinearModelFit would choke on this. However, it looks like I can do a workaround by scaling all of the parameters so that the initial guesses are 1, and then reconvert afterward. All in all, though, this seems particularly non-robust. 5) StepMonitor seems to not work anymore. Do you have any ideas why? $\endgroup$ Aug 19, 2019 at 17:59
  • $\begingroup$ 6) All of this is part of me generating a generic set of routines to do fitting within a particular context. Of the 14 parameters involved, I sort of randomly picked four of them to vary in order to test the code. In a real-world system, many of the parameters can be determined independently, and then fixed to allow the few remaining (hopefully uncoupled) remaining parameters to be determined with more confidence. $\endgroup$ Aug 19, 2019 at 18:01
  • $\begingroup$ (1) @TimLaska gave the definitive answer to your original question. (2) It is not strange for NonlinearModelFit to choke on such fits. The same goes for $R$, $SAS$, MATLAB, Python routines, and every other piece of software that offers optimization routines. (3) Your model essentially has redundant parameters which also makes it harder for any optimization function to work. I'll add to my answer in an attempt to convince you about that. (4) Related to (3) is that the standard errors of the coefficients are way too small because of the parameters redundancy. $\endgroup$
    – JimB
    Aug 20, 2019 at 3:17
  • $\begingroup$ About your item (6): Fixing parameters after estimating them independently is not a recommended statistical procedure. By all means get good initial estimates of all of the parameters that you can. But then the fit needs to be with all parameters estimated simultaneously to get appropriate estimates of predictions and the associated estimates of precision. $\endgroup$
    – JimB
    Aug 20, 2019 at 3:21
5
$\begingroup$

Specifying a constrained global optimization sub method (e.g., NelderMead, DifferentialEvolution, SimulatedAnnealing) to the NMinimize method as described in the tutorial here seems to help with the stability.

Reproducing your code with the NelderMead sub method is shown below.

eqs = {a'[t] == -k1 a[t] - k2 a[t]^2, b'[t] == k1 a[t] + k2 a[t]^2, 
   c[t] == q a[t] + r b[t], c[0] == q a0 + r b0, a[0] == a0, 
   b[0] == b0};
fixedparams = {k1 -> 1.2, b0 -> 0};
fns = {a, b, c};
params = {k2, a0, q, r};
solution = 
  ParametricNDSolve[eqs /. fixedparams, fns, {t, 0, 5}, params];
fitfn = c /. solution;
paramsForDataSet = {k2 -> 1.263, a0 -> 0.0321, q -> 0.341, 
   r -> 0.8431};
dataset = {#, ((fitfn @@ params) /. paramsForDataSet)[#] + 
      RandomVariate[NormalDistribution[0, 0.0002]]} & /@ 
   Range[0, 5, 0.01];
ListPlot[dataset, PlotRange -> Full]
initialGuess = {k2 -> 2.0, a0 -> 0.3, q -> 0.32, r -> 0.88};
tmp = Values@initialGuess;
Dynamic@Column[{Show[ListPlot[dataset, PlotRange -> Full], 
    Plot[(fitfn @@ tmp)[t], {t, 0, 5}, PlotRange -> Full, 
     PlotStyle -> Red], PlotRange -> Full, ImageSize -> Large], 
   ListPlot[{#1, #2 - (fitfn @@ tmp)[#1]} & @@@ dataset, 
    PlotRange -> Full, AspectRatio -> 0.2, ImageSize -> Large]}]
result = NonlinearModelFit[dataset, (fitfn @@ params)[t], 
  Evaluate[List @@@ initialGuess], t, 
  Method -> {NMinimize, Method -> {"NelderMead"}}, 
  StepMonitor :> (tmp = params)]
tmp = Values@result["BestFitParameters"]

NelderMead

$\endgroup$
1
  • $\begingroup$ This is great, thank you! Follow-up question, though. Once I move past my toy system above, I am finding that even THIS solution isn't working. I will update my OP to give an example of that. $\endgroup$ Aug 17, 2019 at 22:32

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.