On Fri, Dec 11, 2009, istvan.kolossvary.hu wrote:
> >We still need to work on this. For simple molecules, TNCG indeed
> >works better
> >than L_BFGS. But for other molecules, TNCG is taking longer than
> >L-BFGS, which
> >is odd. I'll put together some test cases and post them -- maybe there are
> >things we can tweak in the input paramters.
>
> This is the expected behavior. TNCG will always use significantly less
> minimization steps than LBFGS, but a single step can be a lot more
> expensive. It all depends on the Hessian structure of the molecule at
> hand. In terms of total wall-clock time, LBFGS often beats TNCG.
> However, if you need very low gradient RMS, say < 1e-04, TNCG will
> readily beat LBFGS and oftentimes LBFGS would give up before reaching
> convergence.
Sorry: I wasn't being explicit enough: I seem to have examples where
TCNG uses significantly *more* minimization steps (as counted by xmin_iter)
than LBFGS does. Let me put together some test cases, then you can look
at them and see what you think.
...dave
_______________________________________________
AMBER-Developers mailing list
AMBER-Developers.ambermd.org
http://lists.ambermd.org/mailman/listinfo/amber-developers
Received on Fri Dec 11 2009 - 09:00:04 PST
