<html>
<head>
<meta http-equiv="Content-Type" content="text/html; charset=UTF-8">
</head>
<body>
J -<br>
<blockquote type="cite"
cite="mid:202006231951.05NJp7EN057231@ame3.swcp.com">
<meta http-equiv="Content-Type" content="text/html; charset=UTF-8">
<div dir="auto">Is it possible to create agent-based models of
societal collapse? This Nature article argues human societies
have mind- boggling complexity, but I am not so sure if it is
impossible. <br>
</div>
</blockquote>
I appreciate the link to the Nature article... I particularly liked
the caution about misuse of the "Retrospectoscope" , and the
reference to Hari Seldon and PsychoHistory. "... societal spasms
are cyclic" seems nearly tautological, but a good reminder. I
appreciate that many are now treating global
socio-economic-political systems (coupled with the earth-systems) as
a truly complex system as best they can.<br>
<p>Of course it is possible to build an ABM in this domain... bit
as the article gestures toward, there is no bottom to the possible
complexity (at least down to Glen's previous reminders to us that
our individual microbiome is a key part of our "selves" in more
ways than simply accounting for body mass or cell count). <br>
</p>
<p>Unfortunately we aren't just dealing with the risk of our
multiple (highly coupled, but nevertheless diverse) facets of
societies and whole societies collapsing and taking one another
down like dominoes (or a house of cards), but the earth-systems
our high-tech, energy-and-minerals-and-plant-products-hungry
society depends on being able to draw from (exploit?). <br>
</p>
<p>We looked at an SEIR/ABM covid-model for NM at the individual
level which seemed (barely) tractable on a single machine (memory
size)... I have worked with Systems Dynamics models which are
highly aggregative and we even ran 100,000 samples from the World3
model from 1900-2100 as a test/demonstration last year. The
World3 model focuses on Economics and Resource Utilization and was
conceived around the idea of "Limits to Growth"... it doesn't
really do justice to more subtle societal issues (like social
justice, massive political shifts, personal violence, etc.)
Within the 10^5 parameter sweep we started with there are
(naturally) large subregions where human life becomes acutely
miserable. There are lots of criticisms of the model and it IS
very long of tooth. <br>
</p>
<p> I'd be interested if anyone knows of ABMs or Discrete Event
Simulations that aspire to study more than the smallest of
subsets. <br>
</p>
<p>Our experiments with World3 were as much about studying high
dimensional ensemble-problem sets where intuition can be used to
double-check the results, as it was about the problem domain of
impending collapse.<br>
</p>
<div align="center"><img src="cid:part1.57C41999.6819727D@swcp.com"
alt=""></div>
<div align="left">This 2D projection of a 3D "projection" based on
an 11D Self Organizing Map (SOM) was our "best shot" at analyzing
this 100k ensemble. Without stereography or motion parallax, the
3D structure is probably hard to see... it *may* be evident that
one of the most notable features is "continuity" that this appears
to be a complex 2D surface "folded" into 3D. The 2D continuity
is at least partly a direct artifact of our choice of SOM focused
on preserving local distances. The coloring is on a red-blue
spectrum with human population encoding population (ending
population 2100). One of the surprising artifacts found in this
rendering is the relatively regular distribution of high/low
population peppered through the otherwise dominantly low/high
regions. An obvious "positive" correlation between "good
results" and "populations" shows that the sooner we have a
population collapse, the better the chance that the remaining
population will enjoy a high quality of life (by some measure).
We have many hypotheses to test on this rendering, such as whether
the "1D edges" of the embedded surface represent interesting (and
relevant to the problem not our choice of encoding) samples. One
partial result is that (within our sampling) these extrema don't
have any of the "contrary" samples (high among low or
vice-versa)....<br>
</div>
<div align="left"><br>
</div>
<div align="left"> This work is temporarily on hold, but the same
methods are being explored on other problem sets (where there is
funding)... <br>
</div>
<div align="left"><br>
</div>
<div align="left">The main point of this example is that complex
models yield highly complex results and it is not always obvious
or easy to decide what the truly interesting "trade space" is, and
whether or not there are interesting parts of these
high-dimensional landscapes that allow for some useful
intuition. <br>
</div>
<div align="left"><br>
</div>
<div align="left">- Steve<br>
</div>
</body>
</html>