|
What is a Good Agent-based Model?
I can't stop thinking about this business of model validation. Let us still
use professor John Holland's "axiom system". In my mind, I
see three types of "axiom systems".
1. Not much is known about the processes of the system,
but there is an obvious macro pattern. In this case, the modeler can list all
the "axioms" including his assumptions about the mechanism. As long
as his model generates the observed pattern, he can say that the mechanism
postulated could be the true mechanism. It is all right even if he
"manipulates" the 'axioms' as the whole purpose is to explore
possible mechanisms. The ant model and John Holland's language model (how
grammars emerge and how languages evolve) fall into this category. I would
think these models are so called "existence proof model".
2. The model is used to test/explore abstract ideas. The
modeler believes that the system works in a certain way, and he kind of
"proves" his belief (intuition) by generating the pattern observed
in the real world using his model. But the purpose of the model is not to
"prove" his belief, but to say more things about the system.
In this case, it is OK to move all questions of interpretation to
"axioms" and then let the objective program work out, whatever
results are. But there is a limitation to what he says about the system:
whatever points drawn from the model are only claims of the modeler based on
his belief of how the system works. My model and some of the early
social ABMs (like Axelrod's culture dissemination model) fall into this
category. I tend to think these models are more about brain exercise.
3. The model simulates a concrete system and has clear
policy implications (such as Dan Brown et al.'s land-use models). These
models are usually empirically grounded. In this case, only generating the
desired pattern is not sufficient (convincing). And it IS possible to
look into empirical data for realistic and true mechanisms. If the model is
validated at both micro-level and macro-level, the credibility of the model
is enhanced. Of course, no model is a complete representation of the system
and it is not possible to validate an ABM by 100%. But credibility is very
important for this type of models and how well the model is validated much
determines the credibility.
All these three types of models are useful. As they are intended for
different purposes, the requirements for validation are also different. The
judgment of a model depends on its purpose.
|