Page 182 -

9 Verifying and Validating Simulations 179

forward as models of social complexity—can be demonstrated to represent theories
or aspects of social behaviour able to give rise to post-computational models that
are, at some given level, consistent with the onset theories or similar to real data.
Given the model development process described, is there any fundamental
difference between verifying and validating simulations? Rather than being a sharp
difference in kind it is a distinction that results from the computational method.
Whereas verification is focused on the assessment of micro and macro concepts and
inferences in the process of programming, observing and interpreting computational
models, validation is focused on the evaluation of such inferences and concepts as
representations of the target social phenomenon or theory.
In paraphrasing Axelrod (1997a), at first sight, we could say that the problem
is whether an unexpected result is a reflection of the computational model, due to
a mistake in the implementation of the pre-computational model, or is a surprising
consequence of the pre-computational model itself. Unfortunately, the problem is
more complicated than that. In many cases mistakes in the code may not be qualified
simply as mistakes, but only as one interpretation among many others possible
for implementing a conceptual model. Nevertheless, from a practical viewpoint
there may be still good reasons to make the distinction between V&V. A number
of established practices exist for the corresponding quadrants of Fig. 9.1. We will
address some of these in the following sections.

9.3 Validation Approaches

We offered a conceptual definition of validation in Sect. 9.2.2. Had we given an
operational definition, things would have become somewhat problematical. Models
of social complexity are diverse and there is no definitive and guaranteed criterion of
validity. As Amblard et al. (2007) remarked, “validation suggests a reflection on the
intended use of the model in order to be valid, and the interpretation of the results
should be done in relation to that specific context.”
A specific use may be associated with different methodological perspectives
for building the model, with different strategies, types of validity tests, and
techniques (Fig. 9.2). Consider the kind of subjunctive, metaphorical models such
as Schelling’s (1971). In these models there is no salient validation step during the
simulation development process. Design and validation walk together. The intended
use is not to show that the simulation is plausible against a specific context of social
reality but to propose abstract or schematic mechanisms as broad representations of
classes of social phenomena. In other cases, the goal may be modelling a specific
target domain, full of context, with use of empirical data and significant amounts
of rich detail. Whereas in the former a good practice could be modelling with the
greatest parsimony possible so as to have a computational model sanctionable by
human beings and comparable to other models, parsimony can be in opposition to
the goal of descriptive richness and thus inappropriate to the latter case.