AN HYBRID CONNECTIONIST MODEL FOR A DECISION SYSTEM WITH A REJECT OPTION

The reject option, well-known concept in classical pattern recognition methods, is a useful tool for diagnosis problems. Classical multi-layered networks, which are known to approximate a Bayesian classifier, do not allow reject. The purpose of this paper is to propose an original decision system based on neural methods, which solves the reject option problem. The learning process of the network is divided in two differents steps. First we use a competive learning algorithm to create prototypes of the learning set. Then, the system learns a decision using the distances of the patterns to the prototypes. Experimental results in a bi-dimensional and a four dimensional case are shown. They demonstrate the good classification accuracy and the reject ability of the proposed system.