Was there a change related to starting values or stepsize in the latest release in Nelder-Mead fmin? I only saw changes related to convergence criteria in browsing the scipy.optimize history. I'm getting a test failure in statsmodels where the solution has the opposite sign from before. https://github.com/statsmodels/statsmodels/issues/3128#issuecomment-23867326... AFAICS, flipping the sign to negative is still a valid solution. The parameters correspond to the standard deviation and only the square (covariance matrix) is used. However, the starting values are 0.1 and all previous versions of scipy produce positive parameters, i.e. positive standard deviations. The optimization for the test case is pretty deep inside one of the models, and it's not easy to figure out what might be going on. (There are two more statsmodels test failures related to scipy optimize, but I haven't looked at the details yet.) Josef
Tue, 09 Aug 2016 19:03:14 -0400, josef.pktd kirjoitti:
Was there a change related to starting values or stepsize in the latest release in Nelder-Mead fmin? I only saw changes related to convergence criteria in browsing the scipy.optimize history.
I don't think so. Option to specify the initial simplex was also added, but this should not have changed the default one. Indeed, I don't see any change between 0.17.0 and 0.18.0 in examples. If your function modifies the input vector (if yes, probably a bad idea), that may change things, since copying is not necessarily managed the same way.
On Tue, Aug 9, 2016 at 7:27 PM, Pauli Virtanen <pav@iki.fi> wrote:
Tue, 09 Aug 2016 19:03:14 -0400, josef.pktd kirjoitti:
Was there a change related to starting values or stepsize in the latest release in Nelder-Mead fmin? I only saw changes related to convergence criteria in browsing the scipy.optimize history.
I don't think so. Option to specify the initial simplex was also added, but this should not have changed the default one. Indeed, I don't see any change between 0.17.0 and 0.18.0 in examples.
If your function modifies the input vector (if yes, probably a bad idea), that may change things, since copying is not necessarily managed the same way.
Thanks, I consider this as indeterminate sign then, and fix the unit test. I'm not very familiar with that code, but I don't think there is any modification of the input vector. The parameters are put into a larger array that has just a few unkown or unconstrained values. Josef
_______________________________________________ SciPy-Dev mailing list SciPy-Dev@scipy.org https://mail.scipy.org/mailman/listinfo/scipy-dev
participants (2)
-
josef.pktd＠gmail.com -
Pauli Virtanen