Re: Simulating Individual Behavior

[Mark P. Line]

This is my first posting to any of the Swarm mailing lists, so maybe I'm
supposed to introduce myself: Hello, I'm Mark, and I've been fooling
around with microsimulation (and other approaches to what is now called
ABM) off and on since about 1976.

My main interest is in the ability of ABM to unify the scientific
enterprise: different scientific communities talk about the same agents
in different, sometimes contradictory, ways [this is _especially true
in  social and behavioral science], and the explicit realization that
these _are_ the same agents can be the focal point for a deeper level of
consistency among disparate theories and between theory and observation.
That's why this particular thread has interested me a great deal.


Jan Kreft wrote:
> 
> But I can see another way to solve the problem. If you start with a
> conceptual model with all the parameters and mechanics you need and only
> a rough idea of actual values (from population average measurements) you
> can run the sim and compare the output with the real system to be
> modeled. Then you optimize the parameters to achieve the desired output.
> Perhaps call it back-calibration. (Is there an official term for it?)

There is a tradition of (social) microsimulation going back to G. Orcutt
(1957) "A new type of socio-economic system", _Review of Economics and
Statistics_ 58: 116-123 and then the classic Orcutt, G. et al. (1961)
_Microanalysis of Socioeconomic Systems: A Simulation Study_
(NY/Evanston/London).

There's also the more recent Orcutt, G. et al (eds.) (1984)
_Microanalytic Simulation Models to Support Social and Financial Policy_
(Amsterdam).

There's no doubt much more recent stuff, but I've been out of touch with
this particular scene for a while.

The Germans have been doing (social) microsimulation like this in grand
style for about twenty years now, including official applications to
federal policy. Most approaches that I'm aware of distribute parameters
across agents based on dependent frequency distributions that have been
derived from disparate aggregate data sources. I think this may be
equivalent to the kind of "back-calibration" you suggest here, in that
failure to validate a model must necessarily feed back into the
assumptions used to estimate the initial population (and its agents'
parameters). These models never (AFAIK) used anything more
agent-specific than statistical cohorts derived from the dependent
distributions.

Although there is a mountain of literature in German, my favorite is a
published dissertation:

Helmut Vetterle (1986) _Konstruktion und Simulation mikroanalytischer
Modelle. Die Methode der Mikrosimulation und ihre Anwendung_ ("Design
and Simulation of Microanalytical Models: Method and Application of
Microsimulation"). Augsburg: MaroVerlag. >150 refs.

I suspect that this book is out of print, and that there is no English
translation. If I can get Vetterle's consent, I'd be willing to put the
dissertation on my website in the foreseeable future, and even translate
part of it into English -- if there's any interest. (Maybe this is all
old hat and I'm making a fool of myself -- somebody kick me if that's
the case.)


-- Mark

(Mark P. Line  --  Bellevue, Washington  --  <address@hidden>)



                  ==================================
   Swarm-Modelling is for discussion of Simulation and Modelling techniques
   esp. using Swarm.  For list administration needs (esp. [un]subscribing),
   please send a message to <address@hidden> with "help" in the
   body of the message.
                  ==================================