Information Technology Reference
In-Depth Information
patterns of interaction emerge, from which networks might be inferred. In either case,
this raises the possibility of attempting to validate these simulation-generated net-
works against observed real-world networks. Such network validation would be
another way in which to check simulations, and hence increase our confidence in
them. This paper reviews some of such approaches that may be useful for social
simulators.
2
The Central Problem
Networks have a high degree of freedom. There are
2
ways of connecting
n
nodes with undirected edges, even when multiple edges, undirected edges and self-
loops are not allowed. Thus, there are many more ways of joining 25 nodes into such
a network than atoms in the universe (which is around 10
80
, according to [34]). A
single measure projects the set of possible networks onto a single dimension, thus it is
inevitable that a great many different networks will be projected onto small ranges of
the measure. Thus, a lack of divergence of networks with respect to any particular
measure will not be sufficient to distinguish between many distinct networks.
Of course, one rarely wants to prove that one has exactly the same network but ra-
ther a similar one (in some sense); however, this does not change the
scaling
of the
problem. While the tactic of increasing the number of measures used for comparison
may be effective for relatively small networks, it is unlikely to be feasible for 'large'
networks. However, since many permutations of a network exist that are structurally
equivalent [39], this can be used to reduce the number of networks that need to be
considered. One such attempt has been to study the 'universality classes' of the struc-
tural measures/statistics from the network science community but such methods are
rarely addressed in the analysis of agent-based simulated networks [23].
As with any simulation output, not all of its aspects can be considered relevant or
significant with respect to what is being modeled. Some identifiable aspects of the
observed network might be due to the underlying social mechanisms but others mere-
ly due to happenstance (i.e., factors irrelevant to the purpose of the model). Such a
scenario is discussed in [1, 2, 3], where simulated networks from an agent-based
model of friendship networks are compared to a real data of Facebook™ users in a
college campus [1 , 2, 3] . Moreover, not all aspects of the network will have been
intended to be reproduced by the modeler. Thus, when developing agent-based mod-
els that generate (social) networks, it is important to specify:
1.
Which class of networks one might expect to observe if one could “re-run” reality
under the same basic conditions as assumed in the model,
2.
What properties of the simulated networks, one would expect to observe given
how the model has been constructed.
The first comes down to what assumptions one is making about the target social sys-
tem, which (one hopes) is informed by a multitude of empirical sources. The second
depends upon the first but also one's purpose for developing the model and may be
constrained by the available resources.
