MATLAB: Multistart maxtime never exceeded

lsqnonlinmaxtimemultistart

I have been trying to do a nonlinear parameter estimation using lsqnonlin. To get my cost function, I generate measurements, solve an ODE to get my model output and then compute the residuals.

I am using a parallelised MultiStart with 100 different initial points and a matlabpool of size 10. As sometimes my parameters are far from the optimum, ode45 and/or lsqnonlin may take too much time to converge and I decided to set a MaxTime for MultiStart.

The problem is that even when I set a MaxTime of 200s, there are times when it really stops but other times it just hangs indefinitely, even after a walltime of 200s has already gone.

Do you have any idea what could be the problem here? I am running this on a cluster.

Best Answer

I don't know what is going on. But it sounds to me as if your numerical ODE solution might be getting stuck for some values of the parameters, because MultiStart is pretty good about respecting time limits. So I guess that the objective function evaluation is flaky for some parameter values.

Also, have you seen the suggestions in the documentation of optimizing ODEs? There might be something there of interest to you.

Good luck,

Alan Weiss

MATLAB mathematical toolbox documentation

Related Solutions

MATLAB: Intlinprog – MaxTime and LPMaxIter

You cannot set them to inf but you can set them to a large number.

>> opts = optimoptions('intlinprog','MaxTime',inf,'LPMaxIter',inf)
opts = 
intlinprog options:
   Set by user:
    LPMaxIter: Inf
      MaxTime: Inf
   Default:

...

Now whether that's a good idea or not...

MATLAB: Matlab: Get the Hessian matrix using Multistart

MultiStart does not return the Hessian. But, if you really want the Hessian for computing confidence intervals, you shouldn't use the fmincon Hessian anyway, because, as documented, it is inaccurate.

I do not know what the confidence interval calculation is when there are active constraints in a maximum likelihood calculation. If, at the end of the calculation, there are no active constraints, then you can call fminunc on the function at the final point. As documented, fminunc Hessians are accurate.