Statistical learning aims to modelize a functional link between two variables X and Y thanks to a random sample of realizations of the couple (X,Y ). When the variable Y takes a binary number of values, learning is named classification and learn the functional link is equivalent to learn the boundary of a manifold in the feature space of the variable X. In this PhD thesis, we are placed in the context of active learning, i.e. we suppose that learning sample is not random and that we can, thanks to an oracle, generate points for learning the manifold. In the case where the variable Y is continue (regression), previous works show that criterion of low discrepacy to generate learning points is adequat. We show that, surprisingly, this result cannot be transfered to classification talks. In this PhD thesis, we propose the criterion of dispersion for classification problems. This criterion being difficult to realize, we propose a new algorithm to generate low dispersion samples in the unit cube. After a first approximation of the manifold, successive approximations can be realized in order to refine its knowledge. Two methods of sampling are possible : the « selective sampling » which selects points to present to the oracle in a finite set of candidate points, and the « adaptative sampling » which allows to select any point in the feature space of the variable X. The second sampling can be viewed as the infinite limit of the first. Nevertheless, in practice, it is not reasonable to use this method. Then, we propose a new algorithm, based on dispersion criterion, leading both exploration and exploitation to approximate a manifold.
