Background - Design and Analysis of Learning Classifier Systems - page 23

Even though rarely discussed, the necessity of combining the classifier pre-
dictions is an important component of the developed model, as will become
apparent in later chapters. This is a particular property of XCS-like models that
treat their classifiers in some sense independently and thus require their com-
bination at a later stage. For other LCS model types (for example ones that
resemble ZCS), this might not be the case, as will be discussed in the following
chapter.

2.3.5 Michigan-Style vs. Pittsburgh-Style LCS

What has been ignored so far is that there are in fact two distinct types of LCS:
Michigan-style and Pittsburgh-style LCS. In Michigan-style LCS all classifiers
within a population cooperate to collectively provide a solution. Examples are
the first LCS, Cognitive System 1 (CS-1) [116], SCS [95], ZCS [236] and XCS
[237]. In the less common Pittsburgh-style LCS several sets of classifiers compete
against each other to provide a solution with a single fitness value for the set,
with examples for such systems given by LS-1 [198, 199, 200], GALE [151] and
CCS [153, 154].
Even though “Michigan and Pittsburgh systems are really quite different ap-
proaches to learning [ ... ]” [133], they share the common goal of finding sets of
classifiers that provide a solution to the task at hand. Consequently, it is asserted
that their classifier populations can be represented by the same LCS model, but
their way of improving that model is different.
In developing the LCS model we do not distinguish between the two styles, not
even when defining the optimal set of classifiers in Chap. 7, in order to emphasise
that they are just two different implementations that have the same goal. The
point at which this distinction has to be made is as soon as implementation
details will be discussed in Chap. 8.

2.4 Existing Theory

As with the creation of a model for LCS the aim is to also advance the theoretical
understanding of LCS in general, let us review some previous theoretical work in
LCS. Starting with theoretical approaches that consider all LCS subsystems at
once, the focus subsequently shifts to work that concentrates on the GA in LCS,
followed by discussing approaches that have analysed the function approximation
and RL side of LCS.

2.4.1 The Holistic View

The first and currently only LCS model that allows studying the interaction
with the environment and generalisation in the same model was developed by
Holland just after the introduction of the LCS framework [113].
He describes the set of states that the system can take by combining all
possible environmental states and internal states of the

LCS, and defines a