Extracting fuzzy rules from neural networks for land evaluation

In order to make the land evaluation knowledge captured by neural networks transparent, and to make it easy for workers to judge the validity of the land evaluation mode, a method for extracting land evaluation fuzzy rules from a trained neural network was proposed in this paper. Current rule extraction approaches can deal with problems only with discrete-valued inputs or only with continuous-valued inputs. But, the land evaluation factors often contain the continuous-valued, discrete-valued and nominal-valued attributes, a method for constructing neural network with the three kinds of input attributes was proposed firstly. Moreover, an algorithm was developed to extract several main fuzzy rules for each output neuron. The results of experiment illuminate that the proposed method could be utilized to obtain the land evaluation rules from samples that were understandable representation for the users. When the extracted rules does not accord with the fact, the neural networks can be retrained and the rules can be extracted from it again. The proposed method can be more effective to evaluate land than BP neural networks. The validity of land evaluation is improved.