In the text below classifier systems will be introduced. In some sense one can understand multicompartment genomes as classifiers by interpreting one single compartment as an if-then rule. We will discuss here a bit the transformation of a multicompartment genome into a classifier system and how a fitness function can be constructed for such a system. We will do a case study using the simple game 'tic tac toe' as an example (cf. diagram 3.38).
The following points characterize a classifier system compared to a multi-compartment genome:
Independent of the problem to be learned one can state that the set of fitness values must be ordered in a way that it holds for all that . Only then can a system using the fitness values as inputs use these fitness values as a guide to navigate from the 'lower' values to the 'higher' values, thereby 'approaching' the goal states more and more. Because the fitness function is a mapping from the set of possible world states into the set of fitness values one can interpret all those world states which have the same fitness value as 'equivalent' with regard to this fitness value, thus inducing a kind of equivalence classes . From this follows that the set of fitness values should contain at least two different values with .
The challenge for every problem is then to find a fitness function whose domain includes exactly those properties which are important for a LCS system to change its classifiers appropriately.
Gerd Doeben-Henisch 2012-03-31