Geoscience Reference
In-Depth Information
represents your view of the system. It takes time to get rid of semantic bugs; often, you
find the last one after submitting your paper to a journal. The clues that can help: establish
global and/or behavioural parameters in a way that the model works in the most simplistic
and easily predictable way, and check if the model output fits to your expectations. For
example, in the Schelling model set F = 0 or F = 1; in the PARKAGENT model, build a city
of one street, establish one parking place on this street and let one, two and three drivers
enter the city. Try your best to invent as many non-realistic but predictable situations as
possible.
Pass devil's advocate test 1
: Choose minimal and convenient sets of external parameters and
of the parameters of the agent behaviour. Run the model and, in parallel, try to explain the
results by simple calculations, with pen and paper. For example, try to explain why the
Schelling model pattern should converge to a segregated one in the case of f = 0.5. If you
succeeded, repeat with several other sets of parameters. If you repeatedly succeed, aban-
don your simulation and develop an analytic model of the system.
Pass devil's advocate test 2
: Choose a parameter P of interest and divide the range of P varia-
tion into five equal intervals. Run the model five times for the values of P in the middle of
each interval and build dependencies of interesting model outcomes on P. If these depen-
dencies are linear, repeat with other sets of parameters and then add stochastic variation
of the investigated parameter. If all these changes still produce linear dependencies of
the model outcome on the average value of parameter(s), cancel model development and
explain why you get such obvious results.
Pass devil's advocate test 3
: Your model is a complex system in itself and full semantic debug-
ging can never be guaranteed. Carefully test every qualitative effect that you obtain with the
model by varying model settings that seem unimportant for this effect. Vary spatial patterns
of non-animated objects, for example. Never forget that the phenomenon may be artificial,
caused by a specific choice of parameters. It will vanish when parameters are changed.
Model effects should be statistically significant and strong
: Check that the effects that you
consider as meaningful are significant for a reasonable assumption about the standard
deviation of the model parameters. Measure variation of the model outputs using the coef-
ficient of variation (CV = STD × 100%/M): CV above 20% is wonderful, CV ~ 10% is not
bad, CV = 0.5% will hardly cause a reader's reaction.
9.6.4 w
hen
a
PPlying
y
our
M
odel
to
the
r
eal
w
orld
Model validation 1
: If your model has few parameters, there is a chance that you could vali-
date it, that is, compare the model results to real-world data. These data should not be used
for estimating the model parameters. Make this comparison even if you broke the latter
rule, that is, check if your model reproduces the data that you put into it, as sometimes even
this does not happen. Do not forget to mention in your paper that you used the same set of
data to estimate the model parameters and for the validation.
Model validation 2
: You are allowed to choose convenient examples and fail in reproducing
some of the observed phenomena. The model is just a tool for studying reality and it is
quite a success to reproduce part of it. Try to understand and explain the failures.
Every real-world implementation demands essential work:
Be proud of yourself even if you
succeeded in applying your AB model to one real situation only.
9.7 CONCLUSIONS
This chapter presented a detailed overview of AB modelling with a particular emphasis on opera-
tional decisions that an AB modeller must make, for example, the choice of agents, the representa-
tion of relationships between them, the formulation and formalisation of the agent behavioural rules
