A neural architecture for diagnosis

A neural classifier, well-adapted to the diagnosis problem, is presented in this paper. It is able to learn the non-convex envelop of the classes and so, to allow the rejection of points situated far from high density regions in the space. The basic component of the architecture is a Gaussian cell. The learning algorithm allows the incremental recruitment of the cells. A decision rule is proposed and compared to a classical decision rule. Experimental results are shown in a two and four-dimensional case