[MMTK] MMTK 2.0 minimization benchmarks or comparisons?

Konrad Hinsen hinsen@cnrs-orleans.fr
Wed, 15 Dec 1999 10:55:43 +0100


> > > While the steepest descent algorithm is
> > > reasonably fast in MMTK, the conjugate gradients algorithm seems
> > > very slow--I'm guesstimating about 10- to 100-times slower than
> 
> That sounds _very_ wrong. Conjugate gradient should be much faster, more
> accurate, and more reliable than steepest descent. Steepest descent is one
> of those algorithms which should essentially never be used for optimization.

The reason for this observation is that the step counts in the two
implementations mean something very different. For steepest descent,
one step is one energy evaluation plus one displacement along the
energy gradient. For conjugate gradients, one step is the full
minimization along one direction. Therefore one step in conjugate
gradients is much more expensive than one step in steepest descent,
but the number of steps required to reach a minimum is very much
smaller.

Steepest descent is not meant to be used to actually find a minimum,
but it is very useful to get from unrealistic high-energy conformations
to reasonable ones.
-- 
-------------------------------------------------------------------------------
Konrad Hinsen                            | E-Mail: hinsen@cnrs-orleans.fr
Centre de Biophysique Moleculaire (CNRS) | Tel.: +33-2.38.25.55.69
Rue Charles Sadron                       | Fax:  +33-2.38.63.15.17
45071 Orleans Cedex 2                    | Deutsch/Esperanto/English/
France                                   | Nederlands/Francais
-------------------------------------------------------------------------------