Introduction - Design and Analysis of Learning Classifier Systems - page 5

retaining the same techniques to train the model; and if new training methods
for that model type become available, they can be used as a drop-in replacement
for the ones that are currently used.
Clearly, due to the many advantages of the model-based approach, it should
always be preferred to the ad-hoc approach, as the example in this section has
demonstrated.

1.2 Learning Classifier Systems

Learning Classifier Systems are a family of machine learning algorithms that are
usually designed by the ad-hoc approach. Generally, they can be characterised
by handling sequential decision tasks with a rule-based representation and by the
use of evolutionary computation methods (for example, [167, 95]), although some
variants also perform supervised learning (for example, [161]) or unsupervised
learning (for example, [211]), or do not rely on evolutionary computation (for
example, [89]).

1.2.1 A Brief Overview

Based on initial ideas by Holland [109, 110, 111, 109] to handle sequential decision
tasks and to escape the brittleness of expert systems of that time, LCS initially
did not provide the required operational stability that was hoped for [88, 196,
133], until Wilson introduced the simplified versions ZCS [236] and XCS [237],
which solved most of the problems of earlier LCS and caused most of the LCS
community to concentrate on these two systems and their variants.
Learning Classifier Systems are based on a population of rules (also called
the classifiers ) formed by a condition/action pair, that compete and cooperate
to provide the desired solution. In sequential decision tasks, classifiers whose
condition matches the current states are activated and promote their action. One
or several of these classifiers are selected, their promoted action is performed, and
the received reward is assigned to these classifiers, and additionally propagated to
previously active classifiers that also contributed to receiving the current reward.
Occasionally, classifiers of low quality are removed from the current population,
and new ones are induced, with their condition and action based on current
high-quality classifiers. The aim of replacing classifiers is to improve the overall
quality of the classifiers in the population.
Different LCS differ in how they select classifiers, in how they distribute the
reward, in whether they additionally maintain an internal state, and in how they
evaluate the quality of classifiers. The latter is the most significant difference
between early LCS, which based the quality of a classifier on the reward that
it contributed to receiving, and the currently most popular LCS, XCS [237],
that evaluates the quality of a classifier by how accurate it is at predicting its
contribution to the reward.
Shifting from strength-based to accuracy-based LCS also allowed them to be
directly applied to regression tasks [240, 241], which are supervised learning tasks