Can the KA model described in the previous section be somehow implemented? What performances do we expect from it? Why should we prefer such a general model to the many ad-hoc heuristics already existing and performing well in their particular domain?
The first step toward implementation is of course the one of choosing an adequate representation for the knowledge we plan to acquire. It can be espressed both in implicit and in explicit form. It could be, for example, in the form of vectors, or of symbolic expressions, or of neural networks. In a Learning Theory terminology, such a choice defines the hypotheses space.
Independently on the representation, then, a suitable machinery should be provided to ``blindly'' generate conjectures in such a language.
The third element is the one of refutations of wrong theories, or of natural selection. Every conjecture has to be tested, and only the ones which turn out to be ``consistent'' with external stimula survive.
This schema is precisely the one of genetic algorithms [9], [10], with the difference that in Popper model no explicit comment is made about cross-over operator.
Let us review some of the most interesting implementations of such a model in Learning Theory.