This paper proposes a new method for explanation of trained neural networks feed forward type. The new knowledge collected is expressed by fuzzy rules directly from a sensibility analysis inputs/outputs to the neural network. This easy extraction is based on the properties of the derivative of a tangent hyperbolic function used as activation function in the hidden layer of the neural network. The analysis performed is very useful not only for extraction of knowledge, but to know the importance of every rule extracted in the whole knowledge and, also, the importance of every input stimulating the network. An example based on a real case shows the goal properties of the new method proposed.
Keywords: Neural Networks, Neuro-fuzzy models, rule extraction, sensibility analysis, knowledge discovering.
NNA '01, International Conference on Neural Networks and Applications, Tenerife, Canary Islands, pp. 384-391, February 11-15, 2001.
Published: February 2001.