[SciPy-user] advice on stochastic(?) optimisation

David Warde-Farley dwf@cs.toronto....
Thu Aug 28 17:49:10 CDT 2008


On 28-Aug-08, at 11:23 AM, bryan cole wrote:

> I'll looking for a bit of guidance as to what sort of algorithm is  
> most
> appropriate/efficient for finding the local maximum of a function  
> (in 2
> dimensions), where each function evaluation is 1) noisy and 2)
> expensive/slow to evaluate.

Noisy how, exactly? And do you have gradients (or approximate  
gradients)? Can you at least be guaranteed that the function you are  
evaluating is proportional (on average) to the true function?

There is a wide and deep literature on stochastic gradient descent,  
particularly in the context of neural networks. Here are some papers  
that you might find of interest:

Local Gain Adaptation in Stochastic Gradient Descent by N.  
Schraudolph: http://tinyurl.com/69xm45
A set of lecture notes by Leon Bottou on the subject: http:// 
leon.bottou.org/papers/bottou-mlss-2004

In two dimensions, though, I doubt anything too complicated will be  
necessary.

David


More information about the SciPy-user mailing list