relax SVN archive

Added path to other dataset with BFGS.

task #7824(https://gna.org/task/index.php?7824): Model parameter ERROR estimation from Jacobian and Co-variance matrix of dispersion models.

Added dataset which was produced with BFGS.

This shows, that gradient calculation is either wrong in relax, or in minfx.
This is only for minimisation.

It is also easily to see on the graphs.

param: r2a_err, with err: 0.70912495, compared to MC: 0.10395753, with 500 boot 0.66165472
param: dw_err, with err: 0.56449802, compared to MC: 0.03708529, with 500 boot 0.00703459
param: k_AB_err, with err: 0.77161279, compared to MC: 0.29434628, with 500 boot 0.79263287

task #7824(https://gna.org/task/index.php?7824): Model parameter ERROR estimation from Jacobian and Co-variance matrix of dispersion models.

Opened for the possibility to BFGS as minimisation methods.

task #7824(https://gna.org/task/index.php?7824): Model parameter ERROR estimation from Jacobian and Co-variance matrix of dispersion models.

Fix for scaling of parameters in bootstrap.

I cannot find the error, so a relax setup with BFGS will be setup.

task #7824(https://gna.org/task/index.php?7824): Model parameter ERROR estimation from Jacobian and Co-variance matrix of dispersion models.

Added a button for the spectrum.error_analysis user function to the spectra list GUI element.

This is placed after the 'Add' and 'Delete' buttons. The functionality could be improved by
presetting the spectrum ID argument to anything the user has selected in the spectrum list.

Merged revisions 25545 via svnmerge from
svn+ssh://bugman@.../svn/relax/trunk

........
r25545 | bugman | 2014-09-02 14:17:50 +0200 (Tue, 02 Sep 2014) | 8 lines

Another fix for the minfx version checking in the dep_check module.

The version_comparison() function has been created to perform a proper version number comparison by
stripping trailing zeros, converting the two version numbers to lists of int and comparing the lists
using the Python cmp() function. This will return -1 when the version number is too low, 0 when the
versions are equal, and 1 when the version is higher than the minimum.
........

Added scaling to the Jacobian matrix.

But something is very weird with the function values.

Without constraints:
Quasi-Newton BFGS minimisation
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Line search: Backtracking line search.

k: 0 xk: [ 11.8464633, 5.383340723, 10.37739135] fk: 22.1840340996

Parameter values: [11.744468102051812, 5.3843668638102162, 9.8277882468793187]
Function value: 19.089946986005284
Iterations: 8
Function calls: 57
Gradient calls: 9
Hessian calls: 0
Warning: None

With constraints:
Logarithmic barrier function
~~~~~~~~~~~~~~~~~~~~~~~~~~~~
k: 0 xk: [ 11.8464633, 5.383340723, 10.37739135] fk: 73094199769.8
Entering sub-algorithm.

Quasi-Newton BFGS minimisation
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
Line search: Backtracking line search.

k: 0 xk: [ 11.8464633, 5.383340723, 10.37739135] fk: 73094199769.8

Parameter values: [11.846463301422499, 5.3833407234324602, 10.377391345839921]
Function value: 73094199769.763443
Iterations: 1
Function calls: 91
Gradient calls: 2
Hessian calls: 0
Warning: None

task #7824(https://gna.org/task/index.php?7824): Model parameter ERROR estimation from Jacobian and Co-variance matrix of dispersion models.

Another fix for the minfx version checking in the dep_check module.

The version_comparison() function has been created to perform a proper version number comparison by
stripping trailing zeros, converting the two version numbers to lists of int and comparing the lists
using the Python cmp() function. This will return -1 when the version number is too low, 0 when the
versions are equal, and 1 when the version is higher than the minimum.

Initial try to implement using BFGS, when minimising in bootstrap in systemtest: test_task_model_par_est_tsmfk01.

Results:

50 boot

Simplex, with constraints

param: r2a_err, with err: 0.48131663, compared to MC: 0.49354342, with 50 boot 0.45852174
param: dw_err, with err: 0.30874093, compared to MC: 0.32813703, with 50 boot 0.34714120
param: k_AB_err, with err: 0.64606994, compared to MC: 0.65384783, with 50 boot 0.68158598

Simplex, without constraints

param: r2a_err, with err: 0.48131663, compared to MC: 0.49354342, with 50 boot 0.50583526
param: dw_err, with err: 0.30874093, compared to MC: 0.32813703, with 50 boot 0.30275995
param: k_AB_err, with err: 0.64606994, compared to MC: 0.65384783, with 50 boot 0.64242987

BFGS, with constraints
Cannot compute! Extremely show

BFGS, without constraints
param: r2a_err, with err: 0.48131663, compared to MC: 0.49354342, with 50 boot 0.44000111
param: dw_err, with err: 0.30874093, compared to MC: 0.32813703, with 50 boot 0.00554190
param: k_AB_err, with err: 0.64606994, compared to MC: 0.65384783, with 50 boot 0.69600469

Is there something wrong with scaling again?

task #7824(https://gna.org/task/index.php?7824): Model parameter ERROR estimation from Jacobian and Co-variance matrix of dispersion models.

Fix for dimension extractions, when estimating the parameter errors.

task #7824(https://gna.org/task/index.php?7824): Model parameter ERROR estimation from Jacobian and Co-variance matrix of dispersion models.