Re: emergence and parsimony - modeling theory!

[Catherine Dibble]

Sven> Have at it, all you repressed theorists out there!

[irrepressible theorist warning! ;) ]
 
> the practical challenge of emergent behavior is to understand what makes it
> happen, i.e., what makes the micro behavior lead to the observable macro
> behavior, and the relationship between them
>
> the theoretical challenge of emergent behavior is to characterize the
> kinds of relationships that can occur, and to devise ways to compute
> macro from micro more quickly than by using simulation, and to suggest
> choices for or classes of micro from desired or observed macro, to
> reduce the difficulty of inventing or explaining interesting systems

Well put, address@hidden  You've saved me part of my reply, but I also
wanted to add:

Sven> capable of generating a given macro behavior. If we are lucky
Sven> enough to have more than one micro specification that is
Sven> generative in this sense, they suggest we employ Occam's razor
Sven> and go for the 'simplest' one.

cal> finally, we note that parsimony simply does not apply (Occam's
cal> razor is usually the wrong choice for reasoning about biological
cal> systems) - one of the

cal, note that we are not necessarily dealing only with biological
systems, however useful Swarm/ABS may be for modelling them.

Nevertheless, Occam's razor properly suggests "simplest" models only in
the sense that, given a choice, it is usually wisest to choose the
model/explanation that requires the fewest *extraneous* devices.
That's not always quite the same thing as all-around "simplest".

In particular, a good modeller or theorist seeks to simplify as much as
possible *while retaining the essential character* of the system they
seek to model.  I agree with cal that this need not imply the simplest
alternative, but I read Occam's razor somewhat more deeply than that,
and find it a useful rule.  (For example, evolution theory in biology
is a fine example of a parsimonious explanation that allows for exactly
such "historical artifacts".)


Sven> So, should we not also require that our behavior rules at the micro
Sven> level be 'plausible' in some sense? 

I'll argue YES!  In most fields we do have fairly good understanding of
micro-level behavior, whether of chemistry or of economic agents.  As
cal pointed out, Swarm/ABS offer is at least a simulation tool for:
 
cal> "macro" phenomena ... "micro" mechanisms ... so we can hope to
cal> change the one and see a corresponding change in the other, but
cal> there are still not very many useful tools for examining the
cal> consequences of "micro" assumptions
 
That's it exactly, but here's my query re our Swarm/ABS methodology:
What does it mean for us to "see a corresponding change in the other"?

As Fontana and Buss (_Complexity_ 1994, 223-236) point out, "What Would
be Conserved if 'The Tape Were Played Twice'?".  When we are working
with complex systems where no emergent macro pattern is ever exactly
repeated, how do we best classify and *recognize* the meaningful
equivalence classes among our emergent macro phenomena?  (This applies
to our classes of micro behaviours too, but is more direct there.)

I strongly agree with cal that we want to "characterize the kinds of
relationships [between micro and macro] that can occur".  But if we
seek "A generates B" (or at least, "B can be generated by A"!)
relationships within our complex systems, we'll usually need to
establish some detectable essential equivalences among our B outcomes
(since each, by nature of complex systems, will be unique at least at
superficial levels).

Such characterizations are not in themselves our ultimate goal, but
I see them as an essential tool/puzzle along the way to establishing
such relationships.  Although many deeper equivalence classes will
be domain-specific, this still strikes me as one aspect of Swarm/ABS
that is generic and that could benefit from shared insights.  [ideas?]
 
Catherine

*-----------------------------------------------------------------------*
| Catherine Dibble         |                                            |
| Department of Geography  |   "Theory provides the maps that turn an   |
| University of California |    uncoordinated set of experiments or     |
| Santa Barbara, CA 93106  |    computer simulations into a cumulative  |
|                          |    exploration."                           |
|                          |                                            |
| address@hidden       |    -- Booker, Goldberg, Holland (1989)     |
| www.geog.ucsb.edu/~cath  |      _Artificial Intelligence_ 40:235-282  |
*-----------------------------------------------------------------------*