Extending Techniques, Including Implemental Constraints

Abstract

One has actually to solve two sort of problems when devising learning algorithms. Viz., (i) one has to olotain appropriate experimental and, if possible, also theoretical evidence that the procedure in question is really able to learn, it more or less time, the target ⊖ of learning; however (ii) one has also to make efficient use of what is a priori known about the specific properties of the problem. While heuristic ideas concerning these latter properties are usually of help in some in itial period of the learning procedure, one usually needs in the final period a mathematical guarantee that the procedure will really approach to ⊖.