Facing up to the complex problems of stochastic systems as the handwriting recognition, the use ofthe multiple, hybrid and an association of classifier systemsproves an increasing interest during the last years.Basedon complementarities between classifiers, the association of classifiersincreases the performance of the recognition system whilelimiting the error bound to the use of a unique classifier.The use of the multiple classifier systems benefits fromthe strong points of every classifier. Among these systems,we mention the neuro-fuzzy approach. It is about a neuronal approach developed in a fuzzy concept. Our new fuzzy architecture combined the neural fuzzy approach in a hierarchical waywhich offers a solution to the variability of feature vectorsize.In our approach, we use a sequential version of multipleclassifiers. The complementarities between the developedclassifiers are explained in details in talk. Oursystem is based on the use of neural networks developedin a fuzzy concept. The desired outputs of MLPNN are formed using SOM and FKNNA. Therefore,our system is about neuro-fuzzy networks based on SOMand FKNNA association used in the learning process.The first step of the training process of the recognition system consists in the use of theSOM algorithm. This method is applied in order to organize the input prototype vectors. The advantages of thismethod are: This algorithm is a powerful tool for givingthe real classes' number and not the semantic one. Infact, it is an unsupervised algorithm. It makes a projection of the disordered input vectors on an organizedmap of different clusters and it is able to reduce the highdimensionality of our system to two dimensions. Theresult given by this algorithm will be used in the nextstage.In the second step, we adjusted the means of clusterselements, and we calculated the degree of membership ofdata vectors to each cluster by the Fuzzy K-NearestNeighbor Algorithm (FKNNA), which gives a more realistic description. Finally, we attributed the membershipmatrix obtained by FKNNA to the target of the MLPNNin the training process as a desired output of MLPNN.Our experimentation was released using the Arabic digit data, we applied the MLPNN to test the handwrittendigits extracted from MLCA data.Note that the pre-processing system reduces the complexity existing in the principal recognition system by ahierarchical representation. Then, we detail the different sequences of the proposed system of on-line handwritten digit recognition.