Concluding Remarks - Design and Analysis of Learning Classifier Systems - page 232

10 Concluding Remarks
Reflecting back on the aim, let us recall that it was to “develop a formal frame-
work for LCS that lets us design, analyse, and interpret LCS” (see Section 1.3).
Defining LCS in terms of the model structure that they use to model the data
clearly provides a new interpretation to what LCS are. Their design is to be
understood in terms of the algorithms that result from the application of ma-
chine learning methods to train this model in the light of the available data.
Their analysis arises “for free” from the application of those methods and the
knowledge of their theoretical properties.
Regarding the theoretical basis of LCS, most of the existing theory builds on
a facet-wise approach that investigates the properties of sub-components of exis-
ting LCS algorithms by means of representing these components by simplified
models (see Section 2.4). The underlying assumption is that one can gain know-
ledge about the operation of an algorithm by understanding its components.
While one could question if such an approach is also able to adequately capture
the interaction between these components, its main limitation seems to be the
focus on the analysis of existing algorithms, which are always just a means to
an end.
Here, the focus is on the end itself, which is the solution to the problems that
LCS want to solve, and the design of algorithms around it, guided by how LCS
were characterised by previous theoretical investigations. The main novelty of
this work is the methodology of taking a model-centred view to specifying the
structure of LCS and their training. All the rest follows from this approach.
The model-centred view is characterised by first formalising a probabilistic
model that represents a set of classifiers, and then using standard machine lear-
ning methods to find the model that explains the given data best. This results
in a probabilistic model that represents a set of classifiers and makes explicit
the usually implicit assumptions that are made about the data. It also provides
a definition for the optimal set of classifiers that is general in the sense that it
is independent of the representation, suitable for continuous input and output
spaces and hardly dependent on any system parameters, given that the priors
are suciently uninformative. In addition, it bridges the gap between LCS and