[FRIAM] Weighted Ensemble

[Jon Zingale]

Glen,

Yes, that is the kind of weighted ensemble I am thinking about. I see
the algorithm as weighing the novel trajectories more heavily in the
early stages of a search, but ultimately giving the *correct*
stationary distribution in the limit. My list was a combination of
references, dreams, and starting points for a conversation, so thank
you for biting.

>*  2. An alternative (bin-based sampling) to globally defined "fitness" measures in evolutionary modeling.
*

In one paper by Aristoff and Zuckerman, "Optimizing Weighted Ensemble
Sampling of Steady States"[1], the authors mention in passing:

*We emphasize that weighted ensemble is different from most sequential
Monte Carlo methods, as it relies on a bin-based resampling mechanism
rather than a globally defined fitness function like a Gibbs-Boltzmann
potential.*

This caught my attention for a number of reasons. First, it got me
thinking about a trade off between classifying (whether apriori or
adaptive bining) and the globally defined potentials made popular by
physics and saught after by the *fairer* sciences. To the extent that
bining is an alternative implementation of the same underlying algebra
provided by potentials, I got to wondering how I can better see this.
Second, I got to thinking about bining as a kind of niche and how this
approach might correspond nicely to *Natural Design* a'la Thompson and
friends.

>*  3. An application of diffusion-limited aggregation to general search (especially in the face of limited resources)*

So far, here in my early investigations, it appears that questions of
WE optimization and the proving of its low-hanging theorems seem to
rely on the Hill relation[2] and interpreting the process as
diffusion. Here I found visualization of the algorithm at work is
illustrative, the process appears reminescent of watching goatheads[3]
perform a search across my yard (perhaps a kind of diffusion limited
aggreation, DLA, with damping?) Interpreted as a DLA, I can see
*selection* at the bining step as a kind of "collision" mechanism,
causing the DLA to branch. This seems interesting to me because it
helps me to see how one can use a mixuture of maximally-stateful
computations (DLA) and fairly straightforward probabilistic thinking
to perform novel searches. Here, rather than trying to always find the
shorted path or better approximate the mean, WE manages to find
novelty and rare events. idk, it seems promising.

Not mentioned in my previous post, though also of interest, is that
typically weighted ensemble assumes explicit underlying dynamics
(Langevin). But it seems like it should be applicable to networks of
hyperlinks as well. I am thinking of the xkcd comic[4] about
philosophy...

Cheers,

JZ

[1] https://arxiv.org/pdf/1806.00860.pdf

[2] http://statisticalbiophysicsblog.org/?p=8

[3] https://www.google.com/search?q=goathead+plant&oq=goathead+plant&aqs=chrome..69i57.2206j0j7&sourceid=chrome&ie=UTF-8

[4] https://xkcd.com/903/

ps. Since leaving nabble, I haven't figured out how to properly
sub-thread on redfish.com. Any advice/resources are welcome. Also, if
someone out there who can do something is listening, I would like to
be added back onto the emails.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://redfish.com/pipermail/friam_redfish.com/attachments/20210830/37ae2522/attachment.html>