| Popis: |
The paper proposes the concept of Probably Almost Bayesian (PAB) algorithms generalizing a definition given by Valiant in his theory of the learnable. PAB algorithms are defined as algorithms that probably approximate the Bayesian optimum when the training set size tends to infinity, in polynomial time with respect to the training set size. We present this concept in the framework of the decision theory and we support this definition by giving examples of such algorithms, particularly in the field of artificial neural networks. |
