Prev: Direction vector in phase portrait
Next: cross-tabulation, contingency table
From: Frank on 2 Aug 2010 09:41
"Matt J " <mattjacREMOVE(a)THISieee.spam> wrote in message <i36h71$7e9$1(a)fred.mathworks.com>...
> "Frank " <anyone(a)oleole.com> wrote in message <i36fvo$hm1$1(a)fred.mathworks.com>...
>
> > > numerical precision? Ill-conditioning?
> >
> > nope, don't think so.
> >
> > no "strange" options in optimset. only difference in optimset is:
> > LargeScale off no longer necessary in 7.10 but instead I have to set Algorithm to active-set.
> ===================
>
> What I'm asking is, how big is the difference in the solutions? What makes you sure the difference is significant beyond numerical precision?
>
> Also, what makes you certain the minimization problem is numerically stable? When a continuum of (approximate) minimizers exists -- like with the objective function f(x,y)= x^2 where any pair (0,y) is a solution -- small numerical precision differences in the iterative calculations can lead to a different solution along such a continuum.

We are takling about many optimization over a given time horizon T where the solution is a vector x for each t. For most 0<t<T, the difference in each component of x between the two versions around 0.001 however, there are abs(differences) > 0.5 for some t. Given that X(t,i) is around 0.7, a difference of 0.5 is hugh.
From: Matt J on 2 Aug 2010 11:20
"Frank " <anyone(a)oleole.com> wrote in message <i36hu4$n3v$1(a)fred.mathworks.com>...

> We are takling about many optimization over a given time horizon T where the solution is a vector x for each t. For most 0<t<T, the difference in each component of x between the two versions around 0.001 however, there are abs(differences) > 0.5 for some t. Given that X(t,i) is around 0.7, a difference of 0.5 is hugh.
=============

Well, you can check to see if the two different solutions satisfy optimality conditions (e.g. zero gradient, if unconstrained). If they do satisfy optimality conditions, then it's to be expected that fmincon will not distinguish between them. It means you have an issue of numerical stability and need to pose the problem to have a more unique, stable solution.
From: Walter Roberson on 2 Aug 2010 11:22
Matt J wrote:
> "Frank " <anyone(a)oleole.com> wrote in message
> <i36b45$7f3$1(a)fred.mathworks.com>...
>> Dear all
>>
>> I get different results using fmincon under Matlab 7.5 and 7.10. I use
>> exactly the same options.
>>
>> Any known issues?
>
> numerical precision? Ill-conditioning?

Are they being executed on the same computer? If not then differences in
the CPU can lead to differences in result.
From: Frank on 3 Aug 2010 01:43
Walter Roberson <roberson(a)hushmail.com> wrote in message <kpB5o.13018$F%7.4874(a)newsfe10.iad>...
> Matt J wrote:
> > "Frank " <anyone(a)oleole.com> wrote in message
> > <i36b45$7f3$1(a)fred.mathworks.com>...
> >> Dear all
> >>
> >> I get different results using fmincon under Matlab 7.5 and 7.10. I use
> >> exactly the same options.
> >>
> >> Any known issues?
> >
> > numerical precision? Ill-conditioning?
>
> Are they being executed on the same computer? If not then differences in
> the CPU can lead to differences in result.



1. They are executed on the same computer

2. The respective solutions are optimal for the respecitve runs (i.e. solution from 7.5 is optimal for 7.5). However, the solution from 7.5 is not optimal for 7.10 and vice versa.
From: Frank on 3 Aug 2010 02:59
"Frank " <anyone(a)oleole.com> wrote in message <i38a9s$bc8$1(a)fred.mathworks.com>...
> Walter Roberson <roberson(a)hushmail.com> wrote in message <kpB5o.13018$F%7.4874(a)newsfe10.iad>...
> > Matt J wrote:
> > > "Frank " <anyone(a)oleole.com> wrote in message
> > > <i36b45$7f3$1(a)fred.mathworks.com>...
> > >> Dear all
> > >>
> > >> I get different results using fmincon under Matlab 7.5 and 7.10. I use
> > >> exactly the same options.
> > >>
> > >> Any known issues?
> > >
> > > numerical precision? Ill-conditioning?
> >
> > Are they being executed on the same computer? If not then differences in
> > the CPU can lead to differences in result.
>
>
>
> 1. They are executed on the same computer
>
> 2. The respective solutions are optimal for the respecitve runs (i.e. solution from 7.5 is optimal for 7.5). However, the solution from 7.5 is not optimal for 7.10 and vice versa.


one further point:
I have rather accurate starting values. If I use this starting value and do not perform the optimization, all futher output generated by matlab 7.5 and matlab 7.10 is exactly identical. Thus, the only thing that can cause the trouble is the optimization.
First  |  Prev  |  Next  |  Last
Pages: 1 2 3 4
Prev: Direction vector in phase portrait
Next: cross-tabulation, contingency table