glen ropella wrote:
> Swarm is actually very congenial to complex agents. The only hinderance
> is that, since one programs the agents in C, it makes agent design much more
> of a programming task than an agent-design task. I.e. there are no agent
> frameworks to help or hinder you in the design of agents and their internals.
> Effectively, I think this is what you're saying when you say it's more appropriate
> for ALife and SDML is more appropriate for cognitive modeling. Is that right?
Broadly this is right. I would not, however, accept that SDML hinders
you in the
design of agents and their internals. It is true that you cannot build
models in
SDML without agents since it is rulebased and only agents (including the
model) have
rulebases. But then I don't remember that you can build models in Swarm
without
agents.
A difference between Swarm and SDML is that in Swarm you program in C
and glue the
model together with the Swarm libraries whereas in SDML both the
programming of
agents and the "glueing" are done within the same framework. I am a
social scientist
and not a programmer and this integration, together with the debugging
facilities of
SDML, helps me a lot. Speaking personally, I can design, build and
debug models in
SDML in a few days that would take me a couple of weeks in C and Swarm.
Also, I have colleagues who have started to use SDML to prove
propositions about
their models' properties which you can do because SDML is compliant with
strongly
grounded autoepistemic logic.
> If so, then, my question to the list is: Are social scientists mostly
> interested in psychological modeling of social individuals (where some
> cognitive model is assumed), or are they interested in phenomenological
> simulation? Just to clarify, the former has more to do with verification
> of models for understanding how people think. The latter has more to
> do with validation of models for collective, systemic behavior.
>
> What this means to me, a programmer as opposed to a social scientist,
> is that Swarm is good for *finding* models of individual behavior rather
> than assuming one or the other of them, such that those models give rise
> to a collective dynamic. And this is the case even if the models found
> don't fit any prescribed notion of what goes on in the agents "mind".
The problem I have with this approach is the same as the problem I have
with
conventional economists who argue that their models are predicated on
the assumption
that agents act *as if* they were utility or profit maximizers and knew
all of the
opportunities (and dangers) they face as well as every constraint on
their actions at
least up to a subjective (!) probability distribution. How do
economists know that
these are reasonable assumptions? If the model does not yield false
predictions,
they say in Popperian mode. So which economic models never give false
predictions?
None that I know of. How many models have been rejected without further
argument as
disconfirmed because of their false predictions? Again, none that I
know of.I
suppose they would argue that no models always yield correct
predictions. If you are
using them to set policies or take business decisions, surely you will
want to have
some indication of whether the model you will use is likely to give
correct
predictions of the outcomes from your decisions. If you then reject the
assumptions
of the model as descriptions of the conditions in which your model is
applicable,
what is left?
For this reason, I specify a richer agent cognition than is found in
models out of
the SFI but that enables me to use domain expertise (such as decision
makers'
descriptions of how they make decisions and what is important) and also
has some
credibility from experimentally supported theories in cognitive science.
I take my specifications of agent cognition as conditions of application
of my
models. If the results of the models turn out to be incompatible with
observation, I
go back to my domain experts to find out if I have got something wrong.
Sometimes, I
use discrepancies between model outputs to change the representation of
cognition in
order to eliminate those discrepancies and these changes can then be
used as inputs
to institutional analysis by seeing whether the represnetation that
works can be
verified. We thus learn more about the decision making processes
involved.
So my answer to Glen's question is:
I think that good social science entails using model assumptions,
including cognitive
representations, as conditions of application of the models in order to
determine
when models of collective, systemic behaviour are appropriately used for
policy
purposes. If you are not into prediction but use the models to explore
policy
options, then improving the fit between independently validated
representations of
cognition and model outputs informs our understanding of the social
processes, how
they influence individual behaviour and the feedback between the two.
Of course, some models are used to investigate more abstract
propositions about
social relations. The papers on financial markets produced by the SFI
are, to my
mind, superb examples in which abstractly defined, arbitrary cognitive
limitations
are used to investigate the conditions of application of rational
expectations. I
guess that Swarm was much more appropriate than SDML would have been in
that
application because direct application was not an issue and because of
the large
numbers of very long runs involved. Epstein and Axtell's Sugarscape
models would be
better implemented in Swarm. I believe that both are good social science
appropriately supported by Swarm.
I conclude that Swarm and SDML are complements. Perhaps their
developers will learn
from each other.
Finally, mea culpa:
> > Portability is another standard textbook issue in this area. [snip]
> Swarm running on top of Objective C is not portable but running
> > on top of Java is.
>
> I disagree completely with this
> statement; but, it could just be a matter of the defn of "portability".
No, not a matter of definition. I was wrong. I only knew Swarm in
Objective C
running in Unix and I knew noone using it with any other OS. Apologies.
yrs
scott
--
Scott Moss telephone: +44 (0)161 247 3886
Director fax: +44 (0)161 247 6802
Centre for Policy Modelling
Manchester Metropolitan University
Aytoun Building
Manchester M1 3GH
UNITED KINGDOM
http://www.cpm.mmu.ac.uk/~scott
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|
