[Swarm-Modelling] Re: [Swarm-Support] IDEs - who is using what

[Marcus G. Daniels]

Darren Schreiber wrote:
> Game theory and statistics, both have standard ways of doing things 
> that make it substantially easier to figure out what a person has 
> done.  With ABMs, even very skilled programmers tell me about the 
> difficulties they have in replicating results.  [...] I believe this 
> anarchic state and a continuing uncertainty about its epistemological 
> foundations is the reason. 
When I put down a model for a while, and come back to it, I tend to 
forget some of the more delicate but important structures that can drive 
its dynamics.  It's like coming back to working on a complicated program 
or wading into a half-finished paper full of dense notation:  I'm lost 
and senseless until that stuff is back in working memory.

Sure, an obstacle is the technology, and the need to have, to a degree, 
a programming skill set that isn't obviously related to the modeling.  
But, except for truck drivers and pizza delivery people, the skill of 
driving is ancillary to most kinds of work as well.   Driving is a  
skill, you learn it, end of problem.

The more menacing problem with ABMs is how to identify and organize all 
of the pieces so that enough of them get into working memory and one can 
frame feasibly-testable hypotheses about what the simulation should 
under interesting perturbations.  With hundreds or thousands of degrees 
of freedom (either as rules or parameters), the dimensions of an ABM can 
rarely be explored using brute force. 

So I think that's what's needed are not more pretty IDEs and more 
featureful tookits to create ever more complex models (which are ever 
more hopeless to REALLY understand), but instead some usable and 
expressive formal model description techniques.  Like with religion or 
political parties, merely having a standard platform just means that 
people can get accustomed to being confused in the same way.

To have any hope of having a formal model description language 
(especially one where a computer could be used for cognitive support), 
would mean getting away from imperative programming languages like C and 
Java in which side-effects on data are the norm, to functional or 
logic/functional programing languages where they are the exception (e.g. 
Haskell).  To go from "stuff happening", to math..

Questions of computational and analytical tractability will always be 
relevant for agent-based models, but I think verification and validation 
can be made more transparent and automatic given the right formalisms, 
modeler discipline with the formalisms, and software support tools.

Marcus