Hi,
I thought you were up for changing something.
No, as this is your branch. I'll also be on a holiday for two weeks,
starting in a few hours.
There is though one thing there still is a little tricky.
I can't test X, Y values for graphs which are interpolated.
Hmmm, strange!
This could for example be for the model DPL94.
They will change "alot".
The interpolation should not change.
This is because, that the fitted parameters to the model will change,
for each run.
This is again related to the exponential curves, as then the R2eff
errors are determined by Monte
Carlo simulations which are random.
Can you change the test so it uses some other data? Data which
consists of the 2-point approximation rather than the full exponential
curve?
When there is random errors, the chi-square values will also be
subject to random values,
and hence the minimisation will give rise to random results of the
fitted parameters of the model.
R1rho', phi_ex, kex.
Then the exponential curves must be avoided without question in the
test. Alternatively the 'R2eff' model is optimised and the results
saved. These saved results are then loaded by the system test as the
starting values. This saved state trick may not work if the
auto-analysis is involved, but if you run the auto-analysis in the
test you should expect and handle the randomness.
This could be quite confusing for the user.
For every analysis, he/she will get random values and graphs.
Is this a design by purpose?
Methods to solve this, could be:
- Large number of monte carlo simulations, to minimise the difference
in the error between analysis.
A minimum of 500 simulations should be run, as is the default. In
such a case, the parameter variation will be negligible and it should
not confuse the user. A system test cannot afford to run so many
simulations, but a user must do this. It is so fast that they could
run 1000, 5000, or more simulations for higher quality results and
that is why I provide an option in the auto-analysis for specifying
these simulations separately. But 500 is reasonable.
- Make it possible to pass a number to the random generator, thereby
being able to "fix" the randomness.
This is not possible. As mentioned above, for a relaxation dispersion
system test you only have three options available. Use the 2-point
data, use a saved state, or expect and handle the randomness.
Regards,
Edward
Re: [bug #22409] Minimisation of R1rho model give random results