A Learning Classifier Systems Model - Design and Analysis of Learning Classifier Systems - page 43

unchanging number of model parameters while the model is trained, in contrast
to the model structure of a parametric model, which is the part of the model
that is adjusted before training it, and determines the number of adjustable
parameters during model training.
The LCS model that was described in this chapter and forms the basis of
further developments combines a set of local models (that is, the classifiers) to a
global model. While LCS are not parametric models per se, they can be charac-
terised as such by defining the model structure as the number of classifiers and
their localisation, and the model parameters as the parameters of the classifiers
and the ones required for combining the local models. As a result, the task of
training LCS is conceptually split into finding a good model structure, that is,
a good set of classifiers, and training these classifiers with the available training
set.
Finding a good model structure requires us to deal with the topic of model
selection and the trade-off between overfitting and underfitting. As this requires
a good understanding of the LCS model itself, the problem of evaluating the
quality of a model structure will not be handled before Chap. 7. Until then, the
model structure

is assumed to be a constant.
The next chapter discusses how to train an LCS model given a certain model
structure. In other words, it concerns how to adjust the model parameters in the
light of the available data. The temporary aim at this stage is to minimise the
empirical risk. Even though this might lead to overfitting, it still gives valuable
insights into how to train the LCS model, and its underlying assumptions about
the data-generating process. We proceed by formulating a probabilistic model
of LCS in Chap. 4 based on a generalisation of the related Mixtures-of-Experts
model. Furthermore, more details on training the classifiers are given in Chap. 5,
and alternatives for combining the local classifier models to a global model are
given in Chap. 6, assuming that the model structure remains unchanged. After
that we return to developing a principled approach to finding a good set of
classifiers, that is, a good model structure.

M