[SciPy-User] Minimizing Monte Carlo simulation function

[Andrea Gavana]

On 25 September 2014 00:00, Andrew Nelson <andyfaff at gmail.com> wrote:

> >On 24 September 2014 16:07, Matt Newville <newville at cars.uchicago.edu>
> wrote:
> >That said, the comparison above implies that the differential evolution algorithm
> is not better (in either success rate or number of evaluations) than
> basinhopping, but it certainly seems better than anneal was.
>
> I used the benchmarks in the master branch of scipy to obtain the results
> in the gist linked below. The results indicate that deprecating anneal may
> not be a bad idea.
>
> In 9 tests of 'as implemented in scipy'  - the score was:
>
> differential_evolution 6
> basinhopping 3
> anneal 0.
>
> https://gist.github.com/andyfaff/24c96a3d5dbc7b0272b2.
>
> You have to bear in mind that the benchmark functions (contained in
> https://github.com/scipy/scipy/blob/master/scipy/optimize/benchmarks/test_functions.py)
> are quite hard. As such, they aren't approachable on the 'normal'
> minimizers.
>


My own experience with optimizers tells me that DE works very, very well
with specifically designed problems - as you can see from my benchmarks. It
doesn't perform as well on a wider range of benchmarks. It's always the
same story, as I have encountered in the literature countless times: the
functions to be optimized are designed around the algorithm strengths,
Obviously they give the expected results. DE miserably fails for problems
that are non-convex, multi-extrema and with no random behaviour.

Andrea.


http://www.infinity77.net

# ------------------------------------------------------------- #
def ask_mailing_list_support(email):

    if mention_platform_and_version() and include_sample_app():
        send_message(email)
    else:
        install_malware()
        erase_hard_drives()
# ------------------------------------------------------------- #
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mail.scipy.org/pipermail/scipy-user/attachments/20140925/4be13d86/attachment.html>