[FRIAM] Weighted Ensemble

[Roger Critchlow]

This sounds like an algorithm for parallel protein folding that I ranted
about a long time ago.  Start with some collection of conformations;
perform many different molecular dynamics simulations from your starting
points in parallel;  continue with the most promising subset.  As molecular
dynamics on proteins tends to find lots of deadends, you can get a lot of
improvement by tabu'ing the known deadends and extending into conformations
which don't double back into visited regions.  Seems I remember it went
back to some monte-carlo work at LANL in the 1950's, Goodfellow?

It also sounds a lot like Monte Carlo Tree search as used, for instance, in
AlphaGo.

It boils down to how well you can distinguish promising and unpromising
branches.

Whatever, it was in Friam before gmail, so I can't search for it.  There
doesn't appear to be any search in the Friam archives, and the years before
2017 at https://redfish.com/pipermail/friam_redfish.com/ are all 404 anyway.

-- rec --

On Mon, Aug 30, 2021 at 12:39 PM uǝlƃ ☤>$ <gepropella at gmail.com> wrote:

> In my ignorance, I've thought of weighted ensemble (WE) as a specific kind
> of novelty search. E.g. weighting toward trajectories that exhibit
> anomalies. Is that what you mean by it?
>
> Also, for each of the 5 you're interested in, do you have convenient
> example cites for each/any of them? In particular, (2) and (3)? Or are
> these just ideas of places where you think WE should apply?
>
> For my part, no. I haven't used WE in particular. I have a friend who's
> worked on identifying mechanical anomalies from audio (recordings of
> machines as they hum). He may have used it. I'll ask.
>
> On 8/29/21 1:07 PM, Jon Zingale wrote:
> > I am presently working on learning weighted ensemble <
> https://arxiv.org/pdf/1906.00856.pdf> sampling techniques and was curious
> if any here have worked with them before. The technique seems promising and
> has enjoyed quite a bit of success (even above MCMC <
> https://en.wikipedia.org/wiki/Markov_chain_Monte_Carlo>) in circles
> concerned with reaction rates for rare events.
> >
> > Some points of interest for me include:
> >
> >  1. A better sampling of fringe-outlier works/art from streaming
> services.
> >  2. An alternative (bin-based sampling) to globally defined "fitness"
> measures in evolutionary modeling.
> >  3. An application of diffusion-limited aggregation to general search
> (especially in the face of limited resources)
> >  4. An application of linear logic to optimization problems in
> conformation prediction <
> https://en.wikipedia.org/wiki/Protein_structure_prediction>.
> >  5. Investigation of dynamical properties, such as distribution of
> trajectories with "high winding number", on strange attractors.
> >
> >
> > While I am just beginning to grok the technique, I thought it might be
> fruitful to ask here.
> >
> > Jon
>
>
> --
> ☤>$ uǝlƃ
>
> - .... . -..-. . -. -.. -..-. .. ... -..-. .... . .-. .
> FRIAM Applied Complexity Group listserv
> Zoom Fridays 9:30a-12p Mtn GMT-6  bit.ly/virtualfriam
> un/subscribe http://redfish.com/mailman/listinfo/friam_redfish.com
> FRIAM-COMIC http://friam-comic.blogspot.com/
> archives: http://friam.471366.n2.nabble.com/
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://redfish.com/pipermail/friam_redfish.com/attachments/20210830/d4b3e58f/attachment.html>