Complex Adaptive System - Friend or Foe?

[John Sterman]

Content-Type: text/plain; charset="us-ascii"

Niall Palfreyman, in his interesting discussion of SD and CAS modeling
approaches, suggests that "SD typically describes the behaviour of a system
in terms of 5-20 components...."

This is simply not the case. There is a huge range in the size and
complexity of system dynamics models. The simplest teaching models fall in
the range Niall cites. Students at MIT, and, I expect, other schools with
SD programs, routinely build models in their first semester course with
substantially more elements than 20. Most models used in research and
consulting applications are significantly larger, with hundreds or many
thousands of variables. I suspect that the distribution of model sizes is
very broad, so that it is not meaningful to speak of the size of a
"typical" SD model. Such variation in model size arises from the wide
variety of purposes for which models are developed.

The characteristics of a model are determined by the purpose. Discussion
of CAS versus SD in the abstract is therefore misguided and unproductive.
Worse, such discussion reinforces needless conflict about which modeling
methods are "better." There are no sharp demarcations between SD and CAS
models - One can find CAS models with very small numbers of agents and SD
models with large numbers of agents (e.g. firms in an industry, managerial
functions in a firm). Similar variability can be found for other
attributes such as assumptions about the rationality of decision making
assumed for the simulated agents, breadth of model boundary, inclusion of
soft variables, and so on. Models of both "types" can be located at many
points along the many dimensions describing their attributes. Rather than
a divisive debate about whose models are better, let us work together to
build better models, without regard to the labels applied to them for
purposes of marketing or the establishment of a pedigree.

John Sterman
From: John Sterman <jsterman@MIT.EDU>

[Jochen Scholl]

We had a similar discussion three weeks ago (SD3123) under the heading of
"Modeling traditions - flavors of the same thing." The topic is seemingly of
some interest.

I do not see complex adaptive systems modeling (CAS) - also referred to as
agent/individual based modeling(ABM) - as a competing discipline to system
dynamics. Agent-based modelers do not claim to produce superior research,
nor do they attempt to invalidate or replace SD research. They approach
dynamic and complex system problems with a different toolset relying on
different modeling designs (recursive and self-modifying models governed by
only a few rules, chasing for emergent system behavior - "much from little"
as John Holland says).

As long as such basic rules are known, agent-based models can quickly "walk"
through vast model spaces. Ermergent system model behavior can then be
compared with observed phenomena. Models can recursively seek for "better"
rules in order to produce a desired system behavior. The applicability of
such new rules to "real-world" systems can then be studied.

CAS/ABM modeling has, of course, some limitations. If only the emergent
behavior of a system is known from observation, that is, the rules have to
be found which are capable of producing this particular behavior under
observation, then the ABM modeler may be up against a tedious (and maybe
unsolvable) task. In contrast, discovering a dynamic problems underlying
feedback structure as in SD modeling may be less thorny.

There is a rich CAS/ABM literature on many familiar fields of study such as
predator/prey, tragedy of the commons. deer management, and also the beer
distribution game (to mention a few). In a paper listed below, I compared
our SD results with their beer game results as well as those of
"traditional" research. I was amazed to find that the agent-based modeling
results strongly supported our findings while the "traditional" research did
not.

When it comes to choosing the research tools in the study of dynamic and
complex problems, an EITHER-OR approach may preclude us from gaining
additional insights and confidence. At least to me, the opportunity for
triangulation of a dynamic problem seems to be very attractive (when using
both approaches - given their feasibilityalong various dimensions, in the
first place)

CAS/ABM people are increasingly discovering and appreciating SD modeling.
They have similar discussions as we have here. I think it is time to reach
out, compare results when studying the same problems, and understand each
others modeling designs, strengths and limitations more deeply.

(Some of the following I posted earlier this month:) There is a whole
minitrack at the Hawaiian International Conference on System Sciences
dedicated to "Modeling Nonlinear Natural and Human Systems" also covering
the cross-study of Agent-based modeling and System Dynamics.

As a co-chair, I encourage everybody interested in the subject area to
attend and/or contribute to this minitrack (see
http://www.hicss.hawaii.edu/HICSS_35/dtmcfp.htm).

Besides Jim Hiness pioneering work (see
http://web.mit.edu/org-ev/www
esources.htm) one or the other of the
following papers may be of interest:

Miller, J. H. (1998). Active nonlinear tests (ANTs) of complex
simulation models. Management Science, 44(6), 620-830.

Phelan, S. E. (1999). A note on the correspondence between complexity and
systems theory. Systemic Practice and Action Research, 12(3), 237-246.

Scholl, H. J. (2001). Agent-based and system dynamics modeling: a call for
cross study and joint research. Paper presented at the Hawaiian
International Conference on System Sciences, Wailea, HI

Scholl, H. J. (2001, July 23-27). Looking Across the Fence: Comparing
Findings From SD Modeling Efforts With those of Other Modeling Techniques.
Paper submitted to the 2001 Annual International Conference of the System
Dynamics Society, Atlanta, GA.

Kind regards,

Hans J. (Jochen) Scholl
From: Jochen Scholl <JScholl@ctg.albany.edu>

Hans J. (Jochen) Scholl
Center for Technology
in Government/University at Albany/SUNY
phone (518) 442-3892
fax (518) 442-3886
alternate email: JocScholl@aol.com
ICQ 46430532

["Jim Hines"]

>... System Dynamics can not compete with the attractiveness of these
> high-tech state-of-the-art technologies to draw young people..."

Many years ago when doing work in expert systems was incredibly fashionable,
I had a friend who happened to be a Swami in the Kashmir Shavism tradition
and who was also interested in system dyanmics. I observed dejectedly that
work in expert systems was growing "like a weed", and system dynamics was
not. Swami said "Things that grow like weeds usually are weeds."

Regards,
Jim Hines
MIT
jhines@mit.edu

["Jack Ring"]

Ivan,

It is my opinion that if your purposes include discovering, describing
or prescribing systems that do, or must, adapt in behavior and/or
structure and/or content then System Dynamics, the practice, and perhaps
some of the SD-oriented tools even, is complementary to CAS tools such
as Swarm . And even I am proven wrong over time, I think John Sterman
said it well -- it is far too early for anyone to be deciding that an
either/or situation exists.

Although CAS is useful for exploring autocatalytic systems as contrasted
to autopoietic or self-adaptive or regulated systems whereas SD is more
useful for the latter two, it is true that an innovative modeler can
traverse these categories with either practice/tool.

The current pro/con is not about scale or proliferation. It is about;
1) what the modeler has to know, especially about the relationships in
the system, and 2) how much work is involved in articulating the model,
creating the simulation scenarios, making the runs and interpreting the
results.

Regarding what a modeler has to know -- when you proceed to articulate
an SD model you pretty much have to know the nature of the relationships
among the entities. With a CAS model you leverage computer power to
help discover them.

It may be harder to represent autocatalytic behavior with an SD tool but
it can be done. Likewise a simple single-loop feedback regulator can be
modeled in Swarm but not nearly as easily as in iThink or equivalent.

I suggest that the best current "middle ground" is developed and
use-tested by Prof. John Clymer, Cal State Fullerton. He has plug-ins
for the tool Extend that enable modeling of certain types of CAS. This,
coupled with the visual library that Extend now offers lets you leverage
SD thinking with CAS thinking synergistically . And he has written a
385 page manuscript that is both a user guide and courseware.

I think Johns stuff is going to allow me to model the key to all
business and government and society --- the generation, attraction and a
pplication of enthusiasm in an environment of unpredictable change. We
shall see. Has anyone done this in classic SD?

Jack Ring
From: "Jack Ring" <jring@amug.org>

[Cncsdg@aol.com]

Fellow SDers,

I am preparing a presentation to Canadas National Capital System Dynamics
Group next week on Complex Adaptive Systems. It is my opinion that Fractals,
then Chaos and then Complexity and now Complex Adaptive Systems are drawing
the best young researchers and analysts and that System Dynamics can not
compete with the attractiveness of these high-tech state-of-the-art
technologies to draw young people into the field. I was wondering what is
your opinion? Should we embrace CAS or stick to our guns?

Sincerely,
Ivan Taylor
Applications Coordinator
Canadas National Capital System Dynamics Group
cncsdg@aol.com
(613) 995-2445

[Tony Gill]

At 23:22 26/03/01 -0500, Ivan Taylor wrote:
> Should we embrace CAS or stick to our guns?

I think you need to define your modelling purpose before you can begin to
answer such a question. It is perhaps like contrasting algebra with
geometry. They are both useful when correctly applied.

Four areas in looking at this question that immediately come to mind are:
1. Language - Metaphor (CAS) versus explanatory models used to explain
complex behaviour (SD)
CAS is currently best used as metaphor to describe complex behaviour. Sure
there are nice graphs and matrices that can be used to explain behaviour
but these normally are intuitively derived. This comment is based on the
work I have seen to date. (Depends on the rules defined in the limited
range of software.) We all know the power and potential of SD models that
include simulation. CAS perhaps can be contrasted with Causal Loop Diagrams
at the level of language.
2. Modelling approach
CAS tries to model behaviour from the bottom up based on a usually simple
set of rules - ie how do complex systems emerge and adapt. SD models aim to
capture the aggregate behaviour for the system defined. SD models can be
validated. I am not sure this is possible for CAS models.
3. Conceptualisation
I suspect that CAS still has to make significant progress in the area of
conceptualisation which of course influences technology development for
modelling support. Conceptualisation is a well developed part of the SD
method. After all CAS is relatively new when compared to SD.
4. Software
I believe that *at this time* significant CAS modelling is not yet possible
due to the lack of technology - ie, hardware and software (and maybe
brainware see 3 above) are the current limiting factors. SD software is now
well developed to support SD modelling and simulation.

I would welcome comments that differ from the above to help advance my
thinking on this subject.


Tony Gill tel: +44 (0)1295 812262
Phrontis Limited
Beacon House fax: +44 (0)1295 812511
Horn Hill Road
Adderbury email: tonygill@phrontis.com
Banbury
OXON. OX17 3EU web http://www.phrontis.com
U.K.

[Niall Palfreyman]

Cncsdg@aol.com schrieb:
> Should we embrace CAS or stick to our guns?

It may be that Im misunderstanding something somewhere, but I see no
conflict between the two. Complex Adaptive Systems are a family of
dynamical systems which can be modelled, like any other dynamical
system, via stocks and flows. Of course, where they involve spatial
movement of agents, it is certainly difficult to model them in SD, but
otherwise I see no problem in using both tools to complement each other.

Niall Palfreyman.
From: Niall Palfreyman <niall.palfreyman@fh-weihenstephan.de>

[Niall Palfreyman]

Niall Palfreyman schrieb:

> ... otherwise I see no problem in using both tools to complement each other.

Ivan has pointed out to me that Im missing the point of his question. I
said that CAS and SD complement each other, without actually thinking
about _how_ exactly they do so. Which of course is precisely what you
need at this moment Ivan - sorry. The following is my (probably somewhat
naive) perspective on the matter.

I think the complementary nature of the two lies in the fact that CAS
and SD use very different granularities of explanation. CAS describes
the behaviour of a system in terms of thousands (well, in practice
hundreds) of small components. This is a very satisfying thing of
course: out of the mindless behaviour of lots of tiny components arises
something "meaningful" at the macroscopic level. The causes lie on a
significantly finer level of granularity that the effects.

On the other hand, SD typically describes the behaviour of a system in
terms of 5-20 components - it is a far coarser explanation. The causes
lie on approximately the same level as the effects. So doesnt this mean
that SD explanations arent as useful as CAS explanations? No.

The way I see it, the very coarseness of SD explanations forces us to
think very carefully about causes on the same level of granularity as
the effects, and _this_ is the level which most contributes to
understanding. SD is useful _not_ because it tells us the most basic
_physical_ basis of behaviour, but because it tells us the structural
basis of behaviour.

An example: If I want to work out how to make peace between Germany and
England in the 1920s, there are two levels of explanation which might
be helpful to me.
1) I can explain unrest in Germany in terms of the feelings of failure
and loss of pride of individuals as a result of the outcome of the First
World War. This is the CAS explanation; it is at the fine level of
individuals, and offers a possible set of interventions aimed at
creating personal relationships between the two countries. Maybe Ill
encourage student exchange programs, or find other ways to help
individuals in both nations to understand that the people in the other
nation are just that: people. Such interventions are definitely highly
effective in the long-term, but also take a while to implement and are
not easy.

2) I can explain events in the 1920s as interactions between various
European governments; this is the SD explanation - an interaction of
equals, rather than the arising of behaviour out of micro-causes. While
this is in some ways not as satisfying, it is highly effective in terms
of suggesting immediate, and immediately effective, interventions at the
level of governments. England and France are making life financially
extremely difficult for Germany, and may like to consider the fact that
continuing this treatment is not going to help future relations between
the countries. May England will cancel Versaille payments, or arrange a
longer term program which helps Germany through the recession.

Neither of these two kinds of explanation is better than the other. On
the contrary, they together offer a fine- and coarse-level set of
interventions which could effectively complement each other.

The above is straight off the top of my head because I thought it might
be of use to you, Ivan. Its not very well thought out, and may be a
complete load of dingos kidneys. On the other hand it may give you
ideas, and thats the main thing.

All the best,
Niall.
From: Niall Palfreyman <niall.palfreyman@fh-weihenstephan.de>

[Moxnes Erling]

Niall Palfreyman suggests that CAS is better for modelling individual
behaviour and interactions than behaviour of governments (aggregate level
decisions), which is better modelled by SD. Here is a quick response:

This must be correct to the extent that CAS is only particularly useful if
there are more than one individual. With only one individual, I guess CAS
and SD will be identical. However, this does not necessarily imply that SD
is not as useful when there are many individuals. In Nialls example, it is
no problem to include the feelings of failure of individuals in a SD model.
If there is an important feedback loop going through the feelings of
individuals it should be included in the model (as if there is one
aggregate, representative individual in the model). Just as SD is not
limited to systems with long time delays, SD is neither limited to systems
with one or few individuals.

The important question must be if CAS can capture important dynamics and
non-linearities that are not easily captured by SD. The problem focus also
seems important. If a general policy that works on (the average) individual
is sought, SD could be sufficient. If the problem is for instance to
investigate details of a network, CAS could be more appropriate. The
possibility for "bottom up" testing of theories is not limited to CAS, also
SD models typically build on prior information about structure and
parameters.



Erling Moxnes
From: Moxnes Erling <Erling.Moxnes@snf.no>


Erling Moxnes
SNF, Breiviksveien 40, N-5045 Bergen, Norway
Tel. +47 55959526, Fax. +47 55959439

[Yaman Barlas]

The best answer I can come up with is "neither friend nor foe". There is nothing
intrinsic in a method like CAS that would make it friend or foe of sd. The
question is not very well-posed as it tends to justify the unnecessary and
superficial boundaries in between different fields. (But I guess it is somewhat
rhetorical anyway).
But I want to say a few words on the followig part:
"...that System Dynamics can not compete with the attractiveness of these
high-tech state-of-the-art technologies to draw young people..."
While there is some truth to this, I believe it is a "truth" that must be
fought. It is misguided to overemphasize the "high tech" to the extent that the
"technology" becomes a "scientific discipline". If a discipline grows by such
window dressing, it is quite harmful to the field (and to the attracted people)
in the long/medium run. No discipline should be embraced just for its tech
attractiveness. A given CAS research should be embraced if it has a meaningful
and interesting purpose, the research meets the purpose, it is consistent and
well-articulated, etc. The same goes for sd: sd is NOT some simulation software,
stock-flow diagrams or causal-loop diagrams. What makes sd in my opinion a
scientific field is its emphasis on:
- meaningful, realistic, clear equation formulation
- including what is RELEVANT, NOT what is within a given sci. compartment.
- a philosophy that looks for feedback effects that often go unnoticed
- sustainable long term dynamic-patterns, against the short-term event
preoccupation
- Well-articulated, clear, thus TESTABLE models
-...?

The above can be achieved by a huge variety of low and high tech tools -only
some of which are . Can we always meet the above criteria? Often NOT, because
it is not easy. But the crucial point is that the above are in the very
definition of sd. What makes sd a potentially important scientific field are the
above, not the diagrams, eqns, etc. This is unfortunately and precisely what
most critics of sd fail to see. In my view, one of the most important challenges
of the coming decades would be to be able to see and test the "scientific"
contents of the ever increasing new modeling technologies. I am certainly not
saying that CAS is weak in this sense, I do not know. I juts wanted to bring
this issue up in general.
Thanks for the opportunity.
Yaman Barlas
From: yaman barlas <ybarlas@boun.edu.tr>