Page 201 -

198 N. David et al.

9.5.3 Modus Operandi: Formal and Informal Approaches

The tension between simplicity and descriptive richness expresses two different
ways for approaching the construction and validation of a model. One can start
with a rich, complex, realistic description and only simplify it where this turns out
to be possible and irrelevant to the target system—known as the KIDS approach
(Edmonds and Moss 2005). Or one starts from the outset with the simplest possible
description and complexifies it only when it turns out to be necessary to make
the model more realistic (Law 2015), nevertheless keeping the model as simple as
possible—known as the KISS approach (Axelrod 1997a).
In practice, both trends are used for balancing trades-offs between the model’s
descriptive accuracy and the practicality of modelling, according to the purpose and
the context of the model (Sun et al. 2016). This raises yet another methodological
question: the extent to which models ought to be designed on the basis of formal
theories, or ought to be constrained by techniques and approaches just on the
basis of the intuition of the model builders and stakeholders. As we have seen,
strong, subjunctive, ABMs with metaphorical purposes tend to adopt the simplicity
motto with extensive use of formal constructs, making the models more elegant
from a mathematical point of view, easier to verify, but less liable to validation
methods. Game theoretical models, with all their formal and theoretical apparatus,
are a canonical example. Results from these models are strongly constrained by the
formal theoretical framework used.
A similar problem is found when ABMs make use of cognitive architectures
strongly constrained by logic-based formalisms, such as the kind of formalisms
used to specify BDI-type architectures. If the cognitive machinery of the agents
relies on heuristic approaches that have been claimed valid, many researchers in
the literature claim that cognitive ABMs can be validated in the empirical sense of
context-specific models. Cited examples of this kind usually point to ABMs based
on the Soar cognitive architecture (Laird 2012).
At any rate, context-specific models are normally more eclectic and make use
of both formal and informal knowledge, often including informal and stakeholder
evidence in order to build and validate the models. Model design tends to be less
constrained a priori by formal constructs. In principle, one starts with all aspects of
the target domain that are assumed to be relevant and then explores the behaviour
of the model in order to find out if there are aspects that do not prove relevant
for a particular interval of outcomes. The typical approach the majority of all
modelling and validation can be summarised in a cycle with the following iterative
and overlapping steps:
(a) Building and validating pre-computational and computational models: Several
descriptions and specifications are used to build a model, eventually in the
form of a computer program, which are micro-validated against a theoretical
framework and/or empirical knowledge, usually qualitatively. This may include
the individual agents’ interaction mechanisms (rules of behaviour for agents
or organisations of agents), their internal mechanisms (e.g. their cognitive