From: Giacomo Fiorin (giacomo.fiorin_at_temple.edu)
Date: Mon Jan 03 2011 - 18:13:21 CST
Hi Asjasja, the difference is the following:
Let's say that you guessed the set of all the "relevant" collective
variables, which are M in number. For simplicity assume that M = 2, and such
variables are CV(A) and CV(B). In the "multiple-walkers" variant, CV(A) and
CV(B) are the same for all of the replicas, and the biases (i.e. the set of
gaussian hills) are shared by all of them, hopefully speeding up the
sampling by a factor equal to the number of replicas, N. But nothing else
is different from regular metadynamics with just one replica, you still get
a single PMF of the dimensionality that you specified.
In the bias-exchange method, you don't assume to know already which are the
"good variables", nor their number M. You instead specify a larger pool of
them, CV1, CV2, ... CVN, that is supposed to include them (along with other
non-relevant ones). Then you try them all, one at a time or in pairs,
assigned to individual replicas. At regular intervals you swap
configurations between the replicas, thereby sharing the information about
which states get visited. The advantage with respect to regular
metadynamics comes in tough problems where M is much higher than e.g. 2 as
in the example above. Although strictly speaking you don't get an
M-dimensional PMF (which would still require a lot of sampling), you at
least get many 1-dimensional or 2-dimensional PMFs to analyze. Depending on
what you're looking for, these may be good enough.
>From the technical point of view, there's an extra difference. In the
former case you are updating a single PMF, along the few variables that you
chose, and you update it every time a certain replica adds a new hill: but
it's not critical whether the other replicas read such information
immediately. And it is not needed either that all replicas run at the same
speed (meaning NAMD performance). This is good when you are running on a
large supercomputer and some replicas end up on network segments that are
temporarily slower because of higher load.
In the latter case, you swap all replicas at regular intervals, so there has
to be one process that decides the swaps (such as the tclsh process in
replica-exchange NAMD), and all replicas will have to be synchronized at
some point (i.e. the ones running faster will have to wait for the ones
running slower). And you do have to keep track of N different PMFs, instead
of just 1, and find a way to communicate the decided swaps to other
replicas. Finally, another technical issue arises because the Metropolis
criterion should be executed from within the NAMD code, but the
replica-exchange communication goes through the tclsh interpreter outside
NAMD, unfortunately. For these reasons I didn't implement bias-exchange in
NAMD yet, but surely nothing prevents it.
If you absolutely need a bias-exchange implementation already in place, you
may want to look into GROMACS, recompiled with the PLUMED patch.
Dr. Giacomo Fiorin
ICMS - Institute for Computational Molecular Science - Temple University
1900 N 12 th Street, Philadelphia, PA 19122
On Sun, Jan 2, 2011 at 4:03 PM, Ajasja Ljubetič
> I have another question, which is in essence:
> what are the differences between bias-exchange metadynamics and the
> multiple walkers method?
> If I would just run multiple replicas using NAMD, each replica with it's
> own CV bias, isn't this almost a bias exchange method?
> Or must all the replicas share the same CVs? This part of the documentation
> seems to suggest that all the CVs must be the same:
> "*Note: **multidimensional PMFs can only be obtained with one *
> metadynamics* instance applied to all the colvars, and not with multiple
> instances each applied to a single colvar*."
> But I'm not 100% sure I understood this correctly. (also, must the
> potential of the bias for all replicas also be the same?)
> Best regards,
> Ajasja Ljubetič
> On Sat, Jan 1, 2011 at 22:48, Jérôme Hénin <jhenin_at_ifr88.cnrs-mrs.fr>wrote:
>> Dear Ajasja,
>> 2010/12/30 Ajasja Ljubetič <ajasja.ljubetic_at_gmail.com>:
>> > Dear all,
>> > The bias exchange metadynamics is very reminiscent of the replica
>> > method. Since the latter is implemented as a set of TCL scripts, would
>> > be feasible to implement bias exchange in the same way? (ie Does one
>> have to
>> > restart NAMD in order to change the colvar biases?).
>> At this point, yes, you would have to restart NAMD.
This archive was generated by hypermail 2.1.6 : Mon Dec 31 2012 - 23:19:41 CST