On Thu, Aug 06, 2009 at 03:37:25AM -0700, roessli bertrand wrote:
> Hello,
>
> Just have a comment which is also a question concerning dfdp.m
>
> The help states
> "dp= fractional increment of p for numerical derivatives"
> which is fine, but the problem comes that usually nobody knows
> which step to use to calculate the derivatives.
>
> Would it no be better to choose the step in a similar way than it is done
> in __bgsmin.cc?:
>
> SQRT_EPS = sqrt(DBL_EPSILON);
> diff = exp(log(DBL_EPSILON)/3.0);
> // get 1st derivative by central difference
> for (j=0; j<k; j++) {
> p = parameter(j);
> // determine delta for finite differencing
> test = (fabs(p) + SQRT_EPS) * SQRT_EPS > diff;
> if (test) delta = (fabs(p) + SQRT_EPS) * SQRT_EPS;
> else delta = diff;
> // right side
> parameter(j) = p + delta;
> success = __bfgsmin_obj(obj_right, f, f_args, parameter,
> minarg);
> if (!success) error("__numgradient: objective function
> failed, can't compute numeric gradient");
> // left size
> parameter(j) = p - delta;
> success = __bfgsmin_obj(obj_left, f, f_args, parameter,
> minarg);
> if (!success) error("__numgradient: objective function
> failed, can't compute numeric gradient"); parameter(j) = p;
> // restore original parameter for next round
> g(j) = (obj_right - obj_left) / (2.0*delta);
> }
dfdp.m is meant to be called by leasqr.m, which chooses a default of
.001 for dp. This may not be optimal in some cases, but why do you
think the result of the code above is better?
Olaf
------------------------------------------------------------------------------
Let Crystal Reports handle the reporting - Free Crystal Reports 2008 30-Day
trial. Simplify your report design, integration and deployment - and focus on
what you do best, core application coding. Discover what's new with
Crystal Reports now. http://p.sf.net/sfu/bobj-july
_______________________________________________
Octave-dev mailing list
Octave-dev@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/octave-dev
