Re: Optimization using a *Bounded* Levenberg-Marquardt method

Simone wrote:
Hello, I'm trying use the LM method to minimize a function over a bounded
domain. I've looked around and I found some paper regarding the argument. I
have understood the mechanism behind the unbounded LM and I succesfully
wrote a program that works.
The problem is that I don't understand how I must modify my algorithm to add
the fact that I want to optimize over a BOUNDED domain... I mean, from what
I read, I guess I can't simply resize the step lenght if it would bring me
off the domain?
I read about projecting the step vector over my domain... But what if my
domain isn't a vectorial space? I mean, if we work on R^2 and my domain is:
D={(x,y)|x>0,y>0,x+y<1}
I can't project a vector over it..

I hope that my question is understandable and excuse me for my english.

Thanks.

Simone

For x > 0, y > 0, x + y < 1, let
x = e^u/(e^u + e^v + 1)
y = e^v/(e^u + e^v + 1)
and optimize w.r.t. u and v.

.

Relevant Pages
Quantcast