[Top][All Lists]

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

[Help-gsl] Constraints and minimization

From: Max Belushkin
Subject: [Help-gsl] Constraints and minimization
Date: Wed, 30 Nov 2005 18:20:20 +0100
User-agent: Thunderbird 1.5 (X11/20051025)

    Hello everyone,

this post is somewhat related to the thread on multidimensional minimization, but not quite, so I decided to start a clean thread on this one.

I am sure the problem of "constraining some values during minimization" is well known and often encountered, but the issue is only vaguely covered in most articles which employ some methods for this.

Consider the following problem: the function being minimized, f(x), where x is the parameter vector, has several minima. One does not know which one is global, but that is a standard problem. However, the parameter space must be limited: there exist functions g0(x), ..., gN(x), which have upper/lower limits of acceptability. For example, g0<=g(x)<=g1 - i.e. a certain value which is a function of x must lie within the experimental error.

One way of constructing a "constraint" (given no analytic solution is possible) is to construct a dual chi squared function: C = chi squared + constraint. Given chi squared is polynomially bounded, for example, the constraint for "g(x) = g0" can look like this: constraint = pow (P*[g(x)-g0],2)*exp(pow(P*[g(x)-g0],2).

Such a constraint is smooth, and the derivative is smooth, thus it should not spoil the convergence properties of CG methods. Moreover, it contains an adjustable "strength" parameter P, which allows to adjust how strong it is relative to the current chi squared, and what dominates the "total" chi squared fed to the fit.

The problem is, it is far from transparent how it influences each individual parameter in parameter space, and without any possibility of an analytic solution, I do not see how to "map" the constraint over the whole parameter space per parameter and determine which areas of that space will be suitable to constrain each parameter individually, instead of by such a general function which "influences all together, but you don't really know how".

However, this method works for me quite well - convergence without constraints or with constraints does not change significantly.

  Has anyone tried other ways of implementing this?


reply via email to

[Prev in Thread] Current Thread [Next in Thread]