[SciPy-User] Minimizing Monte Carlo simulation function

[Sylvain Corlay]

@Jeremy Besides genetic algorithms, or MCMC methods like SA, there are
other stochastic optimization algorithms which are very well
theoretically founded like Stochastic Gradient Methods
(Robbins-Monro). Stochastic gradient methods actually don't require
knowing the gradient of the objective, but only and "integral
representation" of it, that is, assuming that the gradient has the
form

        DF(x) = E[ H(U,x)]    where U is a random variable and H is a
known function.

Hence, if you are in the context of minimizing a quantity of the form
E[G(U,x)] as a function of x, where G is differentiable wrt x,
stochastic approximation is very often a natural approach and does not
require the problem to be convex. Almost-sure convergence, known rate
of decay of the variance of the error, etc... It is actually rather
simple to implement a ad hoc version of it for your problem. For this
kind of thing, I would avoid pre-canned packages.

Best,
Sylvain

On Fri, Sep 26, 2014 at 4:18 AM, Robert Kern <robert.kern at gmail.com> wrote:
> On Fri, Sep 26, 2014 at 2:59 AM, Sturla Molden <sturla.molden at gmail.com> wrote:
>> On 24/09/14 23:39, Andrea Gavana wrote:
>>
>>> I tend to agree with Matt. Either SA is a weak algorithm or the SciPy
>>> implementation of SA is a weak one. The result doesn't change, it's
>>> still a weak algorithm.
>>
>> It is still useful to have for reference because it is so abundant in
>> the literature. Weakness is a documentation issue.
>
> If you want it around as a reference, please take the code out and
> maintain it in its own project. It's not something that any scipy
> developer wants to maintain, document, or answer questions about. The
> only answer we can give to people is "Don't use it."
>
> --
> Robert Kern
> _______________________________________________
> SciPy-User mailing list
> SciPy-User at scipy.org
> http://mail.scipy.org/mailman/listinfo/scipy-user