Learning criterion and inductive behaviour

摘要

Inductive behaviours may be classified according to their aim. We intend to show that there are at least two kinds of inductive behaviours. Most of the publications seem to take into consideration only one of these: to copy as exactly as possible the behaviour of a probability process. After a brief discussion to explain the necessity of a learning criterion and a recall about one criterion, representative of most of them, we shall define a new criterion, and show why it is better fitted to learn the laws of a deterministic process from a set of observations.This criterion has been used to implement a program which builds an acceptor of natural language sentences in a CAI environment using a tutorial strategy, and then for a question answering device. As attractive as the results are, their improvement requires a semantic model. We give the basic principles of a model which we currently develop, and whose main feature is approximation.