Date: May 28, 2008 12:03 PM
Subject: Re: Preventing negative values in fmincon
John and Bruno,
I think I must be misinterpreting your suggestion. I have
tried several variations of what I think it means, but
nothing seems to work.
Did you mean to take the square root of the variables in the
original objective function without transforming the
variables in the constraints? So because X1 = Y1^2, etc,
then Obj = aX1^0.5 + ... = aY1 + ...
subject to: X1 + ... =< b.
Everything I have tried either stopped because a negative
number was passed into the square root, or has made the
problem less stable. Am I completely missing the point?
"John D'Errico" <firstname.lastname@example.org> wrote in
> "Mike " <michael.gerstREMOVE@yale.edu> wrote in message
> > Bruno and John, thanks for the tip.
> > If I understand you correctly, the following simple LP
> > Obj = a*X1 + a*X2 + a*X3
> > subject to: X1 + X2 + X3 =< b
> > would transform to
> > Obj = a*sqrt(X1^2) + b*sqrt(X2^2) + c*sqrt(X3^2)
> > subject to: sqrt(X1^2) + sqrt(X2^2) + sqrt(X3^2) =< b
> > Is this the basic idea?
> Not exactly. Have the optimizer vary a vecotr
> with three unknowns in it [Y1,Y2,Y3]. Inside
> the objective function, transform your
> variables so that
> X1 = Y1^2
> X2 = Y2^2
> X3 = Y3^2
> So now X1, X2, X3 are assured to be positive.
> When the optimization is done, and it returns
> the vector [Y1, Y2, Y3], just square the elements
> to recover [X1, X2, X3].
> There are similar strategies to deal with variables
> that have both upper and lower bounds.