complexity and society

[Michael Agar]

I'm sorry I missed out on all the interesting threads last week. I  
was swamped helping a center for occupational safety and health in  
California rethink how they design, implement and evaluate their  
training programs using concepts from that thing that we don't know  
what it is (:

I'd like to add a comment.

The first thing is, you don't necessarily need a model to apply  
complexity to society.

If you are considering a model, I like Axelrod's way of thinking  
about them. He sees them as "thought experiment labs" for a  
conclusion based on social research. So first of all the social  
research has to be solid to really do it properly. More often than  
not it isn't.

The lab let's you test arguments of the form, if people do things in  
particular ways properties will emerge at the level of society. By  
"test" I mean it lets you see if the conclusion can be "generated,"  
to use Epstein and Axtell's concept, in just the way your social  
research suggests that it can. It's a way of making the argument that  
underlies the conclusion explicit so it can be better evaluated, and  
it allows for exploration of the space of results that the same  
argument produces and alternative spaces given control parameter  
changes. It's a test of plausibility and an exercise in clarity,  
nothing more, nothing less.

Robert's example of the Iraq war is an interesting one. First of all,  
the question would be, what kind of social research would we do to  
figure out what kind of agent dynamics produce a society that goes to  
war as opposed to one that doesn't. This is an overwhelming project,  
but an important one. Historians and political scientists have  
written much on the subject. Unfortunately there are plenty of recent  
cases we'd have to explore and document as well. We'd  need a large  
number of "real world" runs--i.e. case studies--before we could  
figure out if "virtual" runs were possible to design. Could we even  
simplify enough to conceive of a model? For our many different kinds  
of agents there'd be critical events, diffusion of opinion,  
leadership responses, perceived threat, positions in national  
historical trajectories, all interacting and changing as the story  
developed, with war and peace (to keep it simple) possible emergent  
social system positions that would in turn influence agents and  
perhaps change the trajectory, as is currently going on with Iraq.  
That's just the start of a list of things.

FRIAMers may have already discussed such issues over the week, but I  
thought I'd toss out this particular view of the social/policy model  
theme to help cope with the loss of toothpaste and shampoo on the  
trip home. You should have seen the lines at LAX. But I was rewarded  
with a seatmate who was a professional clown from Albuquerque  
returning from a balloon workshop in Las Vegas. Not your average agent.

Mike

[Jochen Fromm-3]

One question I meet again and again if I try to
make meaningful agent-based simulations is:
- How do we simulate the core of a problem
  without merely constructing an illustration
  of our own beliefs and assumptions ?
In other words: How detailed should an agent-based
simulation be ? If the goal is "to capture the principal
laws behind the exciting variety of new phenomena that become
apparent when the many units of a complex system interact", as
Tamas Vicsek says in http://angel.elte.hu/~vicsek/images/complex.pdf
then how do we design models that are complex enough but not too
complex ?

-If the simulation is too simple and matches your
 own theoretical ideas, then no matter how good these
 ideas are it is always easy to criticize that the
 simulation is either not realistic enough or only
 constructed to illustrate your own ideas and assumptions.
-If the simulation is too complex and matches
 official experimental data, everything takes a
 lot amount of time (creation, setup and execution of
 the experiment and finally the cumbersome analysis
 of the complex outcomes), and it becomes increasingly
 difficult to identify the principal laws, because it is
 easy to get lost in the data or bogged down in details

The "art of agent-based modeling" looks really like an art
to me, something only mastered by a few scientists (for instance
Axelrod). Grimm et al. propose 'pattern-oriented modeling',
Macy and Willer say "Keep it simple" and "Test validity".
What do you think is the best solution for this problem ?

Macy and Willer
"From Factors to Actors: Computational Sociology and Agent-Based Modeling"
http://www.casos.cs.cmu.edu/education/phd/classpapers/Macy_Factors_2001.pdf

Grimm et al.
"Pattern-oriented modeling of agent-based complex systems"
Science Vol. 310. no. 5750 (2005) 987-991
http://www.ufz.de/index.php?de=4976

-J.

-----Original Message-----
From: Michael Agar
Sent: Saturday, August 12, 2006 5:05 PM
To: The Friday Morning Applied Complexity Coffee Group
Subject: [FRIAM] complexity and society

[...] If you are considering a model, I like Axelrod's way of thinking  
about them. He sees them as "thought experiment labs" for a  
conclusion based on social research. So first of all the social  
research has to be solid to really do it properly. More often than  
not it isn't.

The lab let's you test arguments of the form, if people do things in  
particular ways properties will emerge at the level of society. By  
"test" I mean it lets you see if the conclusion can be "generated,"  
to use Epstein and Axtell's concept, in just the way your social  
research suggests that it can. It's a way of making the argument that  
underlies the conclusion explicit so it can be better evaluated, and  
it allows for exploration of the space of results that the same  
argument produces and alternative spaces given control parameter  
changes. It's a test of plausibility and an exercise in clarity,  
nothing more, nothing less. [...]

[Marcus G. Daniels-3]

Jochen,
> -If the simulation is too complex and matches
>  official experimental data, everything takes a
>  lot amount of time (creation, setup and execution of
>  the experiment and finally the cumbersome analysis
>  of the complex outcomes), and it becomes increasingly
>  difficult to identify the principal laws, because it is
>  easy to get lost in the data or bogged down in details
>  
This may be a false choice.   In the case of having some data of
moderate resolution, there's no point in making a hugely elaborate model
and simulation, because you'll never be able to validate beyond your
data anyway.   And if you don't validate, although the modeling still
may be useful as an thought experiment, it isn't science.  You have to
be able to say something that can be shown to be wrong.   If you do aim
to learn things about the world and then predict them it's not desirable
to have giant black box with lots of moving parts.   It's better, if at
all possible, to have a simple story and make the simulation nothing
more than apparatus to help extend the data so that the dynamics can be
studied by theoreticians.

Another mode of use for ABMs is to lower expectations of theoretical
traction and opportunistically look for ways a model makes useful
predictions and then modify the model in that direction over time.  
This is a risky and expensive craft, but one that might have high enough
payoffs to consider (e.g. national security).

It depends on the data and what is of interest.   If the data tells you
about a number of rare events, and it is these events is what you really
care about, then it may make sense to loosely model everyday behaviors
and focus on model microstructure that can create the rare events you
care about.

Finally, sometimes microstructure is known with clearly defined degrees
of freedom, and the dynamics are of interest.  Consider modeling a
factory where different assembly regimes are to be evaluated..  There's
no need to validate here because the whole exercise is to answer
what-ifs about realizable specific systems.

Marcus

[Jochen Fromm-3]

Of course it is the essence of science to verify hypotheses
by experiments. Yet sometimes we have neither suitable
experimental data nor a solid theory, for example
in the case of very large agent-based systems (for instance
for the self-organization and self-management of large
internet applications on planetary scale, or the modeling
of historical processes with millions of actors). It is
hardly possible to examine these systems without simplified
models, and in this case the questions I mentioned seem to
be justified.

In traditional "factor-based" or "equation-based modeling"
we use differential equations and everything is based
on a soild theory: mathematics. This traditional modeling
has a century-long history and we know the suitable parameters,
equations and models. Agent-based modeling has a short history,
we don't know exactly the suitable parameters, agents and models,
and worst of all it is not based on a solid theoretical
theorem-lemma-proof science or calculus like mathematics.

What is missing is a solid science of ABM or a new science of
complexity - something in the direction of Wolfram's NKS idea
(exploring computational universes in a systematic way). Just
as formal, symmetrical and regular systems can be described by
mathematics and 'equation-based modeling', complex systems can
in principle be described by a 'NKS' and agent-based modeling
- which seems to be more an art than a science.

-J.

[Marcus G. Daniels-3]

Quoting Jochen Fromm <fromm at vs.uni-kassel.de>:

> Just as formal, symmetrical and regular systems can be described by
> mathematics and 'equation-based modeling', complex systems can
> in principle be described by a 'NKS' and agent-based modeling
> - which seems to be more an art than a science.

In context, I think is a verification issue.  

ABMs are useful for poking around a complicated system to see what matters and
what doesn't by using a familiar and direct way of describing things, and to
leave the abstractions for later.  ABMs complement traditional techniques of
analysis by extending data.

The imperative programming languages that are typically used to make the
simulations are prone to a variety of programming mistakes but the continue to
be used because 1) they are common and 2) they provide an easy way to think
about side effects (e.g. modifications to a landscape).

Equation-based modelling is more like functional programming, e.g. programming
languages like Haskell that are side-effect free.   I see ABMs moving to these
kinds of programming languages so that components of a simulation can be shown
to be correct, and preferably by automated means.  

As a practical matter, I think it isn't a big deal.  Unit testing during
development by experienced programmers/modelers does a good job of shaking out
bugs.

Marcus

[Michael Agar]

On Aug 14, 2006, at 6:14 AM, Jochen Fromm wrote:

>
>
> - How do we simulate the core of a problem
>   without merely constructing an illustration
>   of our own beliefs and assumptions ?


I'd change this to

How do we make clear the core of a problem through constructing an  
illustration of our own beliefs and assumptions

and say that's exactly what both great science and great art do.  
Science then has the obligation to challenge it against new instances  
of the problem in the classic Popperian way.

Mike

[Jochen Fromm-3]

I think ABMs are more than a tool for poking around in
a complicated system, they can be a tool to understand
complex systems (in the social sciences) and to build
them (in computer science). Every social system is
composed of interacting actors or 'agents'.

In ABMs the word 'model' often means to specify 'the rules
of the game', especially in those areas related to old-fashioned
game theory: to specify exactly what kind of agents exist
(the states) and how they interact (the rules). As Bonabeau
says, "at the simplest level, an agent-based model consists
of a system of agents and the relationships between them".
http://www.pnas.org/cgi/content/full/99/suppl_3/7280

The exploration of these models is a kind of new mathematics,
an exploration of a universe we have defined ourselves. Yet
mathematics is a special science. In science you have to
be able to say something that can be shown to be wrong.
Even if you have no data to verify, or no new theory to
explain something, you can say in mathematics it is true
because you proved it, and anyone can verify if it is
true or not by verifying the proof. What is the equivalent to
axioms, theorems, lemmas and deductive proofs in the NKS of
ABMs ?

[A mathematical theorem has two parts: first a set-up or
set of assumptions (a number of conditions), and second
a conclusion or proposition (a statement which is true under
the given set-up). Does it make sense to say that the first
corresponds to the set-up of an agent-based model, the second
to the associated emergent phenomenon ?]

-----Original Message-----
From: [hidden email]
Sent: Monday, August 14, 2006 5:51 PM
To: The Friday Morning Applied Complexity Coffee Group
Subject: Re: [FRIAM] The art of agent-based modeling

ABMs are useful for poking around a complicated system to see
what matters and what doesn't by using a familiar and direct
way of describing things, and to leave the abstractions for
later.  ABMs complement traditional techniques of analysis
by extending data.

[Marcus G. Daniels-3]

Quoting Jochen Fromm <fromm at vs.uni-kassel.de>:
 
> What is the equivalent to
> axioms, theorems, lemmas and deductive proofs in the NKS of
> ABMs ?

I don't know where it ends, but it must start with the correctness of the
implementation of a simulation (the verification of the model realization).  
In general I don't think they can exist because of issues like the Halting
Problem.  Perhaps recursion theory could say something about when a simulation
is computing something interesting and unattainable in a more compact way...

http://en.wikipedia.org/wiki/Halting_problem
http://en.wikipedia.org/wiki/Recursion_theory

[Jochen Fromm-3]

 
So even if an agent-based model is not realistic enough
to be verified directly by experimental data and if
the simple agents we 'send forth to do battle in our
models' do not produce the same collective behavior as the
apparently real agents, the agent-based model could
serve as a metaphor to understand something. Right ?
Finding new metaphors is indeed something what both great
science and great art do, and since Lakoff's book
"Metaphors We Live By" we know that metaphors are more
than rhetorical elements.

-J.

-----Original Message-----
From: Michael Agar
Sent: Monday, August 14, 2006 5:58 PM
To: The Friday Morning Applied Complexity Coffee Group
Subject: Re: [FRIAM] The art of agent-based modeling

I'd change this to

How do we make clear the core of a problem through constructing an  
illustration of our own beliefs and assumptions

and say that's exactly what both great science and great art do.  
Science then has the obligation to challenge it against new instances  
of the problem in the classic Popperian way.

Mike

[Phil Henshaw-2]

Good methodological questions, except I think it's a mistake, in this
science, to exclude all experiments that don't have a falsifiable
premise.   That's a great methodological 'ratchet' for progress when
it's definable, but a lot of what we're studying with complexity is a
theoretical mess.  I'd even sometimes drop the idea that your research
should even be accumulative, if you're playing with something that
fascinates you.  Mainly though, very good science is done with simple
documentation of observations when people don't quite know what theory
to propose, whether documenting the behavior of artificial or natural
complexity.


Phil Henshaw                       ????.?? ? `?.????
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
680 Ft. Washington Ave
NY NY 10040                      
tel: [phone]                
e-mail: pfh at synapse9.com          
explorations: www.synapse9.com    

> -----Original Message-----
> From: friam-bounces at redfish.com
> [mailto:friam-bounces at redfish.com] On Behalf Of Marcus G. Daniels
> Sent: Monday, August 14, 2006 10:07 AM
> To: The Friday Morning Applied Complexity Coffee Group
> Subject: Re: [FRIAM] The art of agent-based modeling
>
>
> Jochen,
> > -If the simulation is too complex and matches
> >  official experimental data, everything takes a
> >  lot amount of time (creation, setup and execution of
> >  the experiment and finally the cumbersome analysis
> >  of the complex outcomes), and it becomes increasingly
> >  difficult to identify the principal laws, because it is
> >  easy to get lost in the data or bogged down in details
> >  
> This may be a false choice.   In the case of having some data of
> moderate resolution, there's no point in making a hugely
> elaborate model
> and simulation, because you'll never be able to validate beyond your
> data anyway.   And if you don't validate, although the modeling still
> may be useful as an thought experiment, it isn't science.  
> You have to
> be able to say something that can be shown to be wrong.   If
> you do aim
> to learn things about the world and then predict them it's
> not desirable
> to have giant black box with lots of moving parts.   It's
> better, if at
> all possible, to have a simple story and make the simulation nothing
> more than apparatus to help extend the data so that the
> dynamics can be
> studied by theoreticians.
>
> Another mode of use for ABMs is to lower expectations of theoretical
> traction and opportunistically look for ways a model makes useful
> predictions and then modify the model in that direction over time.  
> This is a risky and expensive craft, but one that might have
> high enough
> payoffs to consider (e.g. national security).
>
> It depends on the data and what is of interest.   If the data
> tells you
> about a number of rare events, and it is these events is what
> you really
> care about, then it may make sense to loosely model everyday
> behaviors
> and focus on model microstructure that can create the rare events you
> care about.
>
> Finally, sometimes microstructure is known with clearly
> defined degrees
> of freedom, and the dynamics are of interest.  Consider modeling a
> factory where different assembly regimes are to be
> evaluated..  There's
> no need to validate here because the whole exercise is to answer
> what-ifs about realizable specific systems.
>
> Marcus
>
>
>
> ============================================================
> FRIAM Applied Complexity Group listserv
> Meets Fridays 9a-11:30 at cafe at St. John's College
> lectures, archives, unsubscribe, maps at http://www.friam.org
>
>

[Phil Henshaw-2]

Ah... ok, if you're right!  
But what's the model for finding the gaps in the model?   I think it's
observation, reading the world completely without a model sometimes if
necessary.

> >
> > - How do we simulate the core of a problem
> >   without merely constructing an illustration
> >   of our own beliefs and assumptions ?
>
>
> I'd change this to
>
> How do we make clear the core of a problem through constructing an  
> illustration of our own beliefs and assumptions
>
> and say that's exactly what both great science and great art do.  
> Science then has the obligation to challenge it against new
> instances  
> of the problem in the classic Popperian way.
>
> Mike

Phil
 
> ============================================================
> FRIAM Applied Complexity Group listserv
> Meets Fridays 9a-11:30 at cafe at St. John's College
> lectures, archives, unsubscribe, maps at http://www.friam.org
>
>

[Phil Henshaw-2]

A few posts ago I observed that what's holding back complexity theory
may be not basing it on observation.  One of the things I was thinking
about was the results of the Google search I had just done for "physical
examples of emergence".   I found only two hits!!!!!  One was from a
philosophy book on theoretical biology, relevant enough, and the other
was an artist talking about finding new media for visual expression.
There were no complexity theory hits!  It doesn't say everything, but it
sure says something.

For experiments with ABM's that might explore features of natural
systems, has anyone tried 'composting'?   If emergent structures
decomposed into usable parts would it effect a computational ecology?
Would identifying natural system features and setting up ways to play
with them experimentally be a way to break down the larger task into
workable parts?   Just trying to build whole universes from scratch
seems a rather daunting task.


Phil Henshaw                       ????.?? ? `?.????
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
680 Ft. Washington Ave
NY NY 10040                      
tel: [phone]                
e-mail: sy at synapse9.com          
explorations: www.synapse9.com