The Classifier set 'CLASSIF'

In the examples with the ANIMAT2 structure so far a given classifier set CLASSIF (and later CLASSIF0) is used (cf. 16.3). In a certain sense represents this set of classifiers the memory of the system as well as its knowledge how to proceed in certain situations.

The file zsc2-agent contains two different sets of CLASSIF; only one of these will be used during a real operation.

$\displaystyle c$	$\displaystyle =$	$\displaystyle \langle P, V, A, R\rangle$	(4.85)
$\displaystyle P$	$\displaystyle :=$	$\displaystyle External Sensory Pattern 8 \times 2Bits$	(4.86)
$\displaystyle V$	$\displaystyle :=$	$\displaystyle Internal Sensory Pattern: VITAL$	(4.87)
$\displaystyle A$	$\displaystyle :=$	$\displaystyle Recommended action ACT$	(4.88)
$\displaystyle R$	$\displaystyle :=$	$\displaystyle Cumulated reward REW (through feedbacks)$	(4.89)

In a certain sense one can 'interprete' such an indivdual classifier as a mapping from actual sensory patters from the internal and external world into a recommended mapping

$\displaystyle c_{i}$

$\displaystyle :$

$\displaystyle PERC \longmapsto ACT$

(4.90)

And the whole set of classifiers CLASSIF is then a unification of all these individual mappings forming one big mapping from perceptions into actions. Other interpretations are possible. One can see a classifier as a rule or as a neuron. Therefore an individual classifier has to encode information about possible perceptions as well as possible actions related to these perceptions.

To qualify a classifier as good or bad is only possible in relationship to a goal which is realizable as an environmental setting associated with a certain inner-state setting of the agent itself^4.4. This goal-dependend-dimension has to be encoded as a feedback mechanism. In the case of an emotional classifier system is this realized with the aid of some internal states called emotions (or 'drives' or something else) which function as indicators of the impact of an action mediated through the environment back onto the internal states of the system, encoded as emotions. Presupposing such a body-rooted feedback mechanism it is possible to estbalish a feedback mechanism which operates on a meta level wih regard to the classifiers and by this it is possible to modify the reward (= feedback) values of a classifier with regard to the measured success of the perception-action relationship of the classifier.

In the actual example the set of classifiers is given and constructed manually. In the final version there will be an empty set of classifiers. The system has to find those individual classifiers which are 'helpful' to solve the task which have to be fullfilled in a certain environment with a certain body. This implies that the knowledge of construction is part of such a classifier-building-mechanism. In the realm of biological life this knowledge has been 'collected' during evolution; in our software based theoretical experiments we will assume this knowledge for the next time. Nevertheless these assumptions will be explicit. Everybody can go into this and elaborate an appropriate mechanism to solve this evolutionary.