Page 139 -

136 J.M. Galán et al.

• Rerun the same code in different computers, using different operating systems,
with different pseudorandom number generators. These are most often accessory
assumptions of the executable model that are considered non-significant, so any
detected difference will be a sign of an artefact. If no significant differences are
detected, then we can be confident that the code comprises all the assumptions
that could significantly influence the results. This is a valuable finding that can
be exploited by the programmer (see next activity). As an example, Polhill et
al. (2005) explain that using different compilers can result in the application of
different floating-point arithmetic systems to the simulation run.
Programmer’s activities:
• Reimplement the code in different programming languages. Assuming that the
code contains all the assumptions that can influence the results significantly,
this activity is equivalent to creating alternative representations of the same
executable model. Thus, it can help to detect errors in the implementation.
There are several examples of this type of activity in the literature. Bigbee et
al. (2007) reimplemented Sugarscape (Epstein and Axtell 1996) using MASON.
Xu et al. (2003) implemented one single model in Swarm and Repast. The
reimplementation exercise conducted by Edmonds and Hales (2003) applies here
too.
• Analyse particular cases of the executable model that are mathematically
tractable. Any disparity will be an indication of the presence of errors.
• Apply the simulation model to extreme cases that are perfectly understood
(Gilbert and Terna 2000). Examples of this type of activity would be to run
simulations without agents or with very few agents, explore the behaviour of
the model using extreme parameter values, or model very simple environments.
This activity is common practice in the field.

7.5 Summary

The dynamics of agent-based models are usually so complex that their own
developers do not fully understand how they are generated. This makes it difficult, if
not impossible, to discern whether observed significant results are legitimate logical
implications of the assumptions that the model developer is interested in or whether
they are due to errors or artefacts in the design or implementation of the model.
Errors are mismatches between what the developer believes a model is and what
the model actually is. Artefacts are significant phenomena caused by accessory
assumptions in the model that are (mistakenly) considered non-significant. Errors
and artefacts prevent developers from correctly understanding their simulations.
Furthermore, both errors and artefacts can significantly decrease the validity of a
model, so they are best avoided.
In this chapter we have outlined a general framework that summarises the
process of designing, implementing, and using agent-based models. Using this