Page 39 -

32 P. Davidsson and H. Verhagen

3.4.4 Factors to Consider When Choosing a Model

In contrast to some of the more traditional approaches, such as system dynamics,
individual-based modelling does not yet have any standard procedures that can
support the model development (although some attempts in this direction have been
made, e.g. by Grimm et al. (2006), in the area of ecological systems). In addition,
it is often the case that the only formal description of the model is the actual
programme code. However, it may be useful to use the Unified Modelling Language
(UML) to specify the model.
Some of the modelling decisions are determined by the features of the system
to be simulated, in particular those regarding the interaction model and the
environment model. The hardest design decision is often how the mental state and
the behaviour of individuals should be modelled, in particular when representing
human beings. For simpler animals or machines, a feature vector combined with
a set of transitions rules is often sufficient. Depending on the phenomena being
studied, this may also be adequate when modelling human beings. Gilbert (2006)
provides some guidelines whether a more sophisticated cognitive model is necessary
or not. He states that the most common reason for ignoring other levels is that
the properties of these other levels can be assumed constant and exemplifies this
by studies of markets in equilibrium where the preferences of individual actors
are assumed to remain constant. (Note, however, that this may not always be
true.) Another reason for ignoring other levels, according to Gilbert, is when there
are many alternative processes at the lower level, which could give rise to the
same phenomenon at the macro-level. He illustrates this with the famous study by
Schelling (1971) regarding residential segregation. Although Schelling used a very
crude model of the mental state and behaviour of the individuals, i.e. ignoring the
underlying motivations for household migration, the simulation results were valid
(as the underlying motivations were not relevant for the purpose of Schelling’s
study).
On the other hand, there are many situations where a more sophisticated cognitive
model is useful, in particular when the mental state or behaviour of the individual
constraints or in other ways influences the behaviour at the system level. However,
as Gilbert concludes, the current research is not sufficiently mature in order to give
advice on which cognitive model to use (BDI, Soar, ACT-R, or other). Rather, he
suggests that more pragmatic considerations should guide the selection.
The model of the environment is mostly dictated by the system to be simulated,
with the modeller having to decide on the granularity of the values the environmental
attributes can take. The interaction model is often chosen based on the theory
or practical situation that lies at the heart of the simulation, but sometimes the
limitations of the formal framework used restrict the possibilities. Here, the modeller
also has to decide upon the granularity of attribute values.