Coverage increased (+0.6%) to 76.569% when pullingc5157aa on mp4096:master into7b07af3 on python-control:master.

The original "yottalab" project doesn't exixt anymore. It has been
splitted into two different modules:

• ctrl_repl.py: replacement or implementation of particular functions,
  containing
  • d2c
  • dlqr
  • place (using a wrapper to fortran functions derived from the
    scicoslab project

and

• ctrl_utils with design utilities like:
  • full_obs - full order observer
  • red_obs - reduced order observer
  • comp_form - state feedback controller+observer in compact form
  • comp_form_i - state feedback controller+observer+integ in compact form
  • set_aw - introduce anti-windup into controller
  • grstep - graphical step response

All the other functions of the original yottalab project have migrated
in the control toolbox.

Best regards

Roberto

BTW: For the Swiss Control Association I've prepared a little
description about Python for control purposes.
The link is
http://robertobucher.dti.supsi.ch/bucherdl/BookPythonForControl.pdf

On 07/03/2016 04:36 PM, Mikhail Pak wrote:

Continuing work on the issue raised in#85
#85 and
partially done in#91
#91:

• Replaced most of the |.I|s and |inv|s with |numpy.linalg.solve| or
  |numpy.linalg.lstsq|. The remaining inversions which I didn't
  touch are in |yottalab.py|. Is it still used?
• Where applicable, replaced |A\B|, |A\C| by solving |A[B, C]| and
  taking views for the corresponding results. This avoids redundant
  cubic LU decompositions.
• Added singularity checks before matrix inversions.
• Added two unit tests for the reachable canonical form.
• Refactored an old TODO
  https://github.com/python-control/python-control/blob/master/control/statesp.py#L127.
• Fixed a minor typo: |compatability|
  https://en.wiktionary.org/wiki/compatability -> |compatability|
  https://en.wiktionary.org/wiki/compatibility#English

I commented out the matrix singularity check in |modelsimp.py|
https://github.com/mp4096/python-control/blob/master/control/modelsimp.py#L183
since it correctly raises an error during |testModred()|
https://github.com/python-control/python-control/blob/master/control/tests/matlab_test.py#L364
in |matlab_test.py|. Matrix |A22| is indeed numerically singular in
this case (second column is something like |[[0.0, 3.0e-16]]|). Should
we change the test case?

---

    You can view, comment on, or merge this pull request online at:

#101

    Commit Summary

• Replaced a loop with a map

• Removed redundant numpy.linalg.solve calls

• Numerical improvements in statefbk.py

• Replacedinv withsolve in mateqns

• Added a test case for the reachable canon. form

• Replacedinv withsolve in canonical forms class

• Added a check for unreachable systems to canonical

• Replacedinv withsolve and LSQ in modelsimp

• Fixed "compatability" typo

  File Changes

• M ChangeLog
  https://github.com/python-control/python-control/pull/101/files#diff-0
  (6)

• M control/canonical.py
  https://github.com/python-control/python-control/pull/101/files#diff-1
  (16)

• M control/margins.py
  https://github.com/python-control/python-control/pull/101/files#diff-2
  (2)

• M control/mateqn.py
  https://github.com/python-control/python-control/pull/101/files#diff-3
  (31)

• M control/modelsimp.py
  https://github.com/python-control/python-control/pull/101/files#diff-4
  (30)

• M control/phaseplot.py
  https://github.com/python-control/python-control/pull/101/files#diff-5
  (2)

• M control/statefbk.py
  https://github.com/python-control/python-control/pull/101/files#diff-6
  (7)

• M control/statesp.py
  https://github.com/python-control/python-control/pull/101/files#diff-7
  (20)

• A control/tests/canonical_test.py
  https://github.com/python-control/python-control/pull/101/files#diff-8
  (54)

• M control/tests/mateqn_test.py
  https://github.com/python-control/python-control/pull/101/files#diff-9
  (18)

• M control/xferfcn.py
  https://github.com/python-control/python-control/pull/101/files#diff-10
  (2)

  Patch Links:

• https://github.com/python-control/python-control/pull/101.patch

• https://github.com/python-control/python-control/pull/101.diff

—
You are receiving this because you are subscribed to this thread.
Reply to this email directly, view it on GitHub
#101, or mute
the thread
https://github.com/notifications/unsubscribe/AH9h_sFMOWaLxPBJUZKNeeahe3w7pTNyks5qR8j_gaJpZM4JD4XJ.

Coverage increased (+0.6%) to 76.569% when pulling1fe639c on mp4096:master into7b07af3 on python-control:master.

1fe639c: After some consideration I've also replaced the rank check with bothnumpy.linalg.matrix_rankandnumpy.linalg.det with a singlenumpy.linalg.matrix_rank check. Determinant is really unsuitable for real-life matrix rank computation.

Case in point:

importnumpyasnpA=np.matrix("1.0e-3, 0.0; 0.0, 1.0e-3")print(np.linalg.det(A))print(np.linalg.matrix_rank(A))

This changedoes affect unit tests: Fewer systems are skipped intestAcker.

Coverage increased (+0.5%) to 76.551% when pullingbf6c446 on mp4096:master into7b07af3 on python-control:master.

Coverage increased (+0.5%) to 76.513% when pulling32323d4 on mp4096:master into7b07af3 on python-control:master.

bf6c446: Added singularity check before matrix inversion for DC matching / residualization. I took the liberty to change the unit test slightly, since it does not really matter anyway (see the docstring, it just calls MATLAB-like methods and does not check the results). Instead of eliminating the 2nd and the 3rd state variable, it now eliminates the 1st and the 2nd.

2c25b73: Adds a unit test to check residualization of unstable systems (should throw aValueError exception).

32323d4: Made code more pythonic, in this case list comprehensions instead of for-loops. Also callnp.linalg.eigvals instead ofnp.linalg.eig, for we do not need the eigenvector matrix, only the eigenvalues.

Edit: Spelling

Seehttps://travis-ci.org/mp4096/python-control/builds/144536541 for CI results with the correct configuration from#102.

I will review this tonight.

Coverage increased (+0.3%) to 76.272% when pulling020f64c on mp4096:master into531df81 on python-control:master.

Coverage increased (+0.5%) to 76.1% when pulling71fef4b on mp4096:master into1953399 on python-control:master.

Coverage increased (+0.5%) to 76.1% when pulling71fef4b on mp4096:master into1953399 on python-control:master.

For posterity, the "old TODO" that is referenced in the opening post of this pull request is incontrol/statesp.py at line 127 as of commitc498d3d (current tip ofmaster branch).

The change in your first commit (4d70d00) refactors some code to usemap, whereas changes from PR#110 refactor that code using a list comprehension.

I compared the two alternatives using %timeit of Jupyter for matrices that would correspond to LTI systems with 10 state dimensions, 2 inputs, 2 outputs (so, the A matrix has shape (10, 10), B matrix shape (10,2), etc.), and with 100 state dimensions, 20 inputs, 20 outputs, all using matrices from numpy.random.random((r,c)).

The difference in timing performance does not appear significant. However, the case ofmap was slightly faster (approximately 500 to 900 ns less in duration), so I recommend that we use it.

@mp4096 Thanks for these improvements of numerical reliability.

As future work, I think that it would be interesting to add more test cases that would fail to be detected ifnumpy.linalg.inv was used instead ofnumpy.linalg.solve.