Geoscience Reference
In-Depth Information
Computers, Environment and Urban Systems
(Guo and Mennis, 2009) for more papers on this
subject. When a collective pattern can be recognised during a simulation, an agent's reaction to the
emerging pattern can then be considered as a component of the behavioural rules. For this reason,
my advice to AB modellers is to carefully check the performance of the data mining procedures
used and to invest in fast methods of pattern and cluster recognition.
9.2.7 a
S
with
e
Very
M
odel
, aBM, e
Ventually
, r
eacheS
the
S
tage
of
V
alidation
Students of natural science are brought up in an atmosphere of validation: any model of real-world
phenomena remains just an intellectual exercise until it is validated. Numerous aspects of valida-
tion express different views regarding the comparison of a model's forecast to the real phenom-
enon. Here, I just mention, according to Sargent (2013), the main stages of validation as commonly
accepted in the recent literature on the subject:
Face validation
is a kind of a Turing test (Turing, 1950): Experts (usually yourself) watch the
behaviour of the agents in the model and decide if it is reasonable.
Historical data validation
: This is what physicists call validation - use some of the data for
estimating model parameters and compare the remaining data with the model outputs.
Parameter variability - sensitivity analysis
: Vary the input and internal parameters of a
model and estimate the effect of this variation on the model's dynamics.
Predictive validation
: Forecast the system's dynamics and compare with real data. This lux-
ury is rarely possible in geographic AB models.
An excellent review of validation approaches with regard to AB modelling was undertaken by Ngo
and See (2012), who also present a comprehensive list of references.
9.2.8 c
hoice
of
the
S
oftware
e
nVironMent
Should an AB modeller use specialised AB software? The question is especially important when
you make your first AB experiments, and the answer depends on your programming experience.
The experienced programmer does not need my advice. To my students who usually (1) are not
afraid of installing and activating Microsoft Visual Studio or some Java environment, (2) can code
and execute the Game of Life in an hour+ and (3) do not need a tutor for writing macros in Excel,
I suggest starting with NetLogo. NetLogo's critical advantage is fast prototyping. Its programming
language is easy to learn, the model library contains numerous examples of simple models and con-
venient controls enable you to build your own version of the Game of Life during the first 2 h. Erez
Hatna's web version of the Schelling model, referred to in Section 9.1.7, is an excellent example of
what can be done with NetLogo.
The simplicity of the environment turns into a disadvantage when you aim to build a larger
model, such as the PARKAGENT model described previously. The NetLogo programming lan-
guage is a procedural and not an object-oriented, programming environment, which is intentionally
thin, differs greatly from the standards of Java, C# or C++, and is not able to communicate with
databases. All these disadvantages will become important in a couple of months, once you have
spent time altering the behavioural rules of the agents every other morning. After you roughly
decide on the agents, the relationships, attributes and, most importantly, the behavioural rules of
the agents, reconsider your choice of programming environment. Recent reviews of AB modelling
software will help you to choose between specialised programming environments and common
programming languages (Crooks and Castle, 2012).
Let us now illustrate the process of AB model development and investigation with the Schelling
and PARKAGENT models. I have already used these models for presenting general aspects of AB
modelling given earlier.
