Introduction - Design and Analysis of Learning Classifier Systems - page 7

developed within other machine learning methods, but now such a position is
dicult to argue for.

1.3 About the Model-Centred Approach to LCS

This work arises from the lack of theoretical understanding of LCS,andthe
missing formality when developing them. Its objective is to develop a formal
framework for LCS that lets us design, analyse, and interpret LCS.Inthat
process it focuses on related machine learning approaches and techniques to
gain from their understanding and their relation to LCS.
The immediate aim of this work is not to develop a new LCS.Ratheritisto
give a different perspective on LCS, to increase the understanding and perfor-
mance of current LCS, and to lay the foundations for a more formal approach to
developing new LCS. Neither is the introduced model to be taken as the LCS
model. It was chosen for demonstrative purposes, due to its similarity to the
popular XCS. Other LCS model types can be constructed and analysed by the
same approach, to represent other LCS types, such as ZCS.

1.3.1 The Initial Approach

The initial approach was to concentrate on an LCS structure similar to XCSF
[240] and to split it conceptually into its function approximation, reinforcement
learning and classifier replacement component. Each of these was to be analysed
separately but with subsequent integration in mind, and resulted in some studies
[78, 83, 155] for the function approximation component and others [79, 80, 81]
for the reinforcement learning component.
When analysing these components, the goal-centred approach was followed
both pragmatically and successfully: firstly, a formal definition of what is to be
learned was given, followed by applying methods from machine learning that
reach that goal. The algorithms resulting from this approach are equivalent or
improve over those of XCSF, with the additional gain of having a goal definition,
a derivation of the method from first principles, and a strong link to associated
machine learning methods from which their theoretical analysis was borrowed.
When concentrating on classifier replacement, however, taking this approach
was hindered by the lack of a formal definition of what set of classifiers the pro-
cess of classifier replacement should aim at. Even though some studies aimed at
defining the optimal set for limited classifier representations [130, 133, 135], the
was still no general definition available. But without having a formally expressible
definition of the goal it was impossible to define a method that reaches it.

1.3.2 Taking a Model-Centred View

The definition of the optimal set of classifiers is at the core of LCS:givena
certain problem, most LCS aim at finding the set of classifiers that provides the
most compact competent solution to the problem.