Introduction - Design and Analysis of Learning Classifier Systems - page 9

1.3.4 Novelties

The main novelty of this work are a new methodology for the design and analysis
of LCS, a probabilistic model of their structure that reveals their underlying
assumptions, a formal definition of when they perform optimally, new approaches
to their analysis, and strong links to other machine learning methods that have
not been available before.
The methodology is based on taking the model-centred approach to describing
the model underlying LCS, and applying standard machine learning methods
to train it. It supports the development of new LCS by modifying their model
and adjusting the training methods such that they conform to the new model
structure. Thus, the introduced approach, if widely adopted, will ensure a formal
as well as empirical comparability between approaches. In that sense, it defines
a reusable framework for the development of LCS.

1.4 How to Read This Topic

Many concepts that are frequently used in this work are introduced throughout
the text whenever they are required. Therefore, this work is best read sequen-
tially, in the order that the chapters are presented. However, this might not be
an option for all readers, and so some chapters will be emphasised that might
be of particular interest for people with a background in LCS and/or ML.
Anyone new to both LCS and ML might want to first do some introductory
reading on LCS (for example, [42, 133]) and ML (for example, [19, 102]) be-
fore reading this work from cover to cover. LCS workers who are particularly
interested in the definition of the optimal set of classifiers should concentrate on
Chapters 3 and 4 for the LCS model, Chapter 7 for its Bayesian formulation
and the optimality criterion, and Chapter 8 for its application. Those who want
to know how the introduced model relates to currently used LCS should read
Chapters 3 and 4 for the definition of the model, Chapters 5 and 6 for trai-
ning the classifiers and how they are combined, and Chapter 9 for reinforcement
learning with LCS. People who know ML and are most interested in the LCS
model itself should concentrate on the second half of Chapter 3, Chapter 4, and
Chapter 7 for its Bayesian formulation.

1.4.1 Chapter Overview

Chapter 2 gives an overview of the initial LCS idea, the general LCS framework,
and the problems of early LCS. It also describes how the role of classifiers
changed with the introduction of XCS, and how this influences the struc-
ture of the LCS model. As our objective is also to advance the theoretical
understanding of LCS, the chapter gives a brief introduction to previous
attempts that analyse the inner workings of LCS and compares them with
the approach that is taken here.
Chapter 3 begins with a formal definition of the problem types, interleaved with
what it means to build a model to handle these problems. It then gives a