[SciPy-Dev] scipy.optimize.anneal - deprecation

[Andrew Nelson]

On 13 October 2014 10:37, Andrew Nelson <andyfaff at gmail.com> wrote:

> On 13 October 201
>>
>> Looking at the tests at
>> http://infinity77.net/global_optimization/multidimensional.html  the
>> scipy version of simulated annealing, SIMANN, performs horribly. However,
>> ANA seems to do pretty well. So the problem with scipy seems to have been a
>> poor version of the algorithm and probably we should just fix that.
>>
>> Chuck
>>
>
> Andrea Gavana kindly made the code to the benchmark suite available.  I'm
> modifying to a form for inclusion into scipy:
> https://github.com/andyfaff/scipy/tree/go_benchmark/scipy/optimize/benchmarks.
> As we speak I'm running the test optimization functions (I added a few and
> modified some) for basinhopping, differential_evolution and anneal.
> I'll put these up in a Gist in the next day or so.
>
>

As promised, the first run through of the test functions for global
optimizers is at: https://gist.github.com/andyfaff/24c96a3d5dbc7b0272b2.
This was for a total of 150 random starting vectors.  There are still some
things to be ironed out, particularly how a success is judged (atol vs
rtol, etc).  For example, in the Thurber problem (a NIST regression
standard) a lot of failures are because the minimizer used for polishing
(L-BFGS-B) doesn't seem to be able to take the energy from 5645 to 5642.
The parameters are so close to the optimum solution I'm wondering if it's a
precision problem with the numerical derivatives.  I think most of those
fails could be counted as successes.

In any case, I'm not going to comment any further, but will let the numbers
tell their own story. For each problem the percentage of successes is
reported (whether the minimizer found the global minimum), as well as the
mean number of function evaluations (irrespective of whether it was a
success or failure).

regards,
Andrew.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.python.org/pipermail/scipy-dev/attachments/20141013/47306ebc/attachment.html>