Réseaux de neurones récurrents à fonctions de base radiales: RRFR - Application au pronostic

This paper introduces a Recurrent Radial Basis Function network (RRBF) for non-linear system prognosis. The training process is divided in two stages. First, the parameters of the RRBF are determined by the unsupervised k-means algorithm. The ineffectiveness of this algorithm is improved by the FuzzyMinMax technique. In the second stage, a multivariable linear regression supervised learning technique is used to determine the weights of the connections between the hidden and output layer. We test the RRBF on the Box and Jenkins furnace database. This application shows that the RRBF is able to predict the evolution of a non-linear system. The performances of the RRBF are compared with those of the TDRBF. The RRBF gives better results for long run predictions. The FuzzyMinMax technique makes the K-means more stable.