Genetic Reorganization

With the preceding mechanisms it is possible to reinforce existing classifiers and to generate new ones. Nevertheless it can happen that the resulting set of classifiers is not well prepared to enable a good overall performance. Thus the set of all classifiers $CLASSIF$ represents the description of a function $f_{C}$ which has to be optimized. It is known that the application of genetic operators to such a function $f_{C}$ can optimize this function. But the question is, how often such a genetic modification (optimization) should happen? A minimal frequency could be to apply every time when a positive goal has been explicitly met. Otherwise, if a certain set of classifiers $C_{i}$ is 'bad' and eventually would not allow to find a goal at all it would be 'wise' to enable a reorganization more often, e.g. every $n$-many cycles. Thus one has to determine some value for $n$ which is 'sound'^4.2.

1. Select two classifiers $\{c_{i}, c_{j}\} \subseteq CLASSIF$ based on collected reward REW
2. Copy $\{c_{i}, c_{j}\}$ and generate offsprings $\{c'_{i}, c'_{j}\}$
3. Apply crossover or mutation to the perception $C[e,...]$ and to the action $C[...,A,...]$ part of a classifier following fixed probabilities yielding $\{c^{+}_{i}, c^{+}_{j}\}$. The reward part $C[...,R]$ of each offspring is the the mean of the rewards of the parents.
4. Insert the offsprings $\{c^{+}_{i}, c^{+}_{j}\}$ into $CLASSIF$
5. If the amount of $CLASSIF \vert CLASSIF\vert$ shall be fixed to a certain size one has to delete two other classifiers $\{c_{r}, c_{s}\}$ whose total reward is below '0'.