Learning Classifier Systems (LCS)

Figure 12.1: GA realized within a classifier system

Learning Classifier Systems (LCSs)(cf. 12.1) are a kind of a specialization of the general concept of a GA because within LCSs the population of 'genes' is formatted as a population of classifiers. A classifier is a unit combining three parts: an IF-part -also called condition-, an ACT-part representing some kind of an action', and a F-part, representing some kind of fitness values, which are feeded into the classifier either from the outside or from the 'carrier system' - the 'body' - of the system, which interfaces the information with 'semantical' structures. The the set of classifiers can be called population (POP) (a subset of some bigger set $P$ of all possible classifiers) as in the case of the genetic algorithms. But in this context this set has a different functional role; it is a population of entitites bound to one system and is functioning for this system as a kind of knowledge $KNOW$ influencing the observable behavior $\Phi_{OBS}$ of the system as a special kind of internal states $KNOW \subseteq IS$:

$$\Phi_{TH} : I \times KNOW \times (IS-KNOW) \longmapsto MATCH$$ (12.1)

where a input string $\sigma \in I$ is matched with the IF-parts of the systems knowledge $KNOW$. Those classifiers which match are then candidates for the generation of a $MATCH$ set with $MATCH \subseteq KNOW$. The final selection of an action depends from fitness states $F_{SYS}$ which are part of the knowledge $KNOW$.

In the classical versions of LCS systems (see 'Some Bits of History' below) the fitness values have been deduced from the environment without including a learning system SYS (often also called 'agent'). Therefore was the LCS theory somehow unfinished or incomplete. A major factor of the learning process was somehow 'outside' of the scope of the theory. Everybody who did use an LCS system had to provide it's 'own' model of fitness generation.

In recent years one can observe a shift in the way how to look to the fitness values. In Neuropsychology e.g. (c.f. Solms et al.(2002)[] one could work out that the body itself has several basic processes managing survival which communicate with the brain and which are signaling built-in fitness values for a basic orientation. And this new concept of 'built-in fitness values' has already found its way into robotics (see e.g. the paper Gordon et al. (2010)[123]). But there were several precursors to Gordon and sometimes the label of emotional robotics has been used too.