Re: Integration problem

[fredomatic]

Hi!
The solution to too much accuracy required is to INCREASE the tolerance
to errors, not to decrease it. Relative and absolute tolerances of 1E-6
are usually easy to meet for most problems, but you can go to 1E-4, for
example. I would not go much higher than that, because you risk
instability. You can also play with the stiffness flag, and you could
also try the Cvodes integrator, but I would fist check that, with
average parameter values, a single simple simulation works. I would also
check the model mass balance (or similar conservation requirement),
because integration problems often happen when some state variables grow
out of hand. Finally, I would check whether the prior distributions
assigned generate unrealistically high or low values, because that leads
to difficulties in integration. In that case, it is better to use
truncated distributions.
Best regards.
Frederic

On 13/04/2021 16:19, Bastian Hirn-Derksen wrote:
> Dear all,
> 
>  
> 
> I hope someone of you can help me with a problem in GNU McSim I obtain,
> when I wont to calibrate my model via MCMC. When I start MCMC
> calibration Im getting the following error message :
> 
> Doing 10000 Metropolis within Gibbs simulations
> lsodes-- at start of problem, too much accuracy   
>          requested for precision of machine..  see tolsf (=r1)
>  in above message, r1=      1.#QNAN00000000
> Warning: DoOneExperiment failed
>
> 
> Im aware that there must be an issue with the ode solver and the
> integration of the data, but Internet research did not provide any
> solution. I also tried to set the atol and rtol value to zero for having
> absolute error tolerance but this also did not work and occured in this
> error message:
> 
> Doing 10000 Metropolis within Gibbs simulations 
> lsodes-- at start of problem, too much accuracy  
>          requested for precision of machine..  see tolsf (=r1)
>  in above message, r1=      1.#QNAN00000000
> Warning: DoOneExperiment failed
> 
> 
> I hope you can help me with this problem. Calibration of the model is
> essential for my work and therefore I have to find a solution of the
> problem. 
> 
> With kind regards,
> 
> --
> 
> Bastian Hirn-Derksen
> 
> Apotheker
> 
> Westfälische Wilhelms-Universität Münster Institut für Pharmazeutische
> und Medizinische Chemie
> 
> - Klinische Pharmazie -
> 
> Raum A 120.202
> 
> Corrensstraße 48
> 
> 48149 Münster, Germany
> 
> Phone: +49 251 83 33364
> 
> Fax.: +49 251 83 32144
> 
> E-mail: _b_hirn01@uni-muenster.de <mailto:b_hirn01@uni-muenster.de>_
> 
>  
>