RC failing with line Search

[AlexAv]

Hi

I am running different models with E-Surge. To improve convergence and precision of the estimates, I changed the tolerance to 10-e8 and I allowed for 3 iteration cycles.
But a the end of the iterations, I often get the message:
"RC failing with Linesearch". What does this message mean?
This message doesn't appear when the tolerance is set to its default values for the same model.

In addition, sometimes, the convergence drops a the end of the iteration process (I can see that on the DOS window) while before it seems to have reached a minimum (e.g: the deviance is minimal and seems to be stable from the 20th to the 30th iteration, and the it drops to infinite, until the end of the optimization, i.e; until the last 200th iteration). Again, what does that mean? Does that mean that I reach a global minimum and then the deviance drops because the space around the minimum is explored too? or does that simply mean that I reached some of the potential local minima and hence that I can't really trust the estimates.

if someone has already experienced that

thanks in advance

Sincerly

Alex

[CHOQUET]

Hello Alexy,

concerning

> I am running different models with E-Surge. To improve convergence and precision of the estimates, I
> changed the tolerance to 10-e8 and I allowed for 3 iteration cycles.
> But a the end of the iterations, I often get the message:
> "RC failing with Linesearch". What does this message mean?
> This message doesn't appear when the tolerance is set to its default values for the same model.

This means that the solver didn't succeed to improve the convergence. A tolerance of 10-e8 is very low because this value is close to the square root of the machine precision. This is the best that can be acheived by non-linear solvers.

> In addition, sometimes, the convergence drops a the end of the iteration process (I can see that on
> the DOS window) while before it seems to have reached a minimum (e.g: the deviance is minimal and > seems to be stable from the 20th to the 30th iteration, and the it drops to infinite, until the end of
> the optimization, i.e; until the last 200th iteration). Again, what does that mean? Does that mean
> that I reach a global minimum and then the deviance drops because the space around the minimum is > explored too? or does that simply mean that I reached some of the potential local minima and hence
> that I can't really trust the estimates.

This think can happens when some individuals histories (a few of them) are quite different from the others. In that case when the solution is close to an optimum for most of the individual histories,
the probabilities of these histories become close to zero, because the model doesn't fit well these histories.

Rémi

[AlexAv]

Hi Rémi,

thanks for your reply.
So, if I understand well, we don't really know if convergence is attained or not. Actually the message 'says' that the solver can't do more, is that true?

Futhermore, I am running the models with initial values (IV) from previous selected models to improve convergence.
for some models, I reduced the numbers of groups in order to reduce the time needed for building constraints using gemaco (for example if the groups code for sex, year and age factors, but if the model only includes a sex effect, I will only used two groups) . In that the case, I can use the IVs, and I can run the model.
However, for more complex models, involving for example a year effect (and so using up to 27 groups), I can get the constraints with Gemaco, but I can't fix the IVs in the IVFV box (I waited for three days, and I have never reached the run step). Should I wait longer time?

again, thanks in advance,

Alex

[CHOQUET]

> So, if I understand well, we don't really know if convergence is attained or not. Actually the message 'says' > that the solver can't do more, is that true?

Yes, you cannot do better that the machine precision. The second point is that some individual capture
histories are certainly not compatible with your model.

The time in IVFV is spent to built a reduced parameter set of mathematical parameter.
There is a option in E-SURGE which allow you to disconnect the building of this set.