Extraction of fuzzy rules using neural networks with structure level adaptation and its application to diagnosis for hepatobiliary disorders.

ABSTRACT

First, this paper presents the reasoning and the learning method for fuzzy rules using structure level adaptation of neural networks. In a usual neural network's mechanism, during learning process of rules, we can observe the following two behaviors Case 1 If a neural network does not have enough neurons to be satisfied to infer, then the input weight vector will have a tendency to fluctuate greatly, even after a certain long period the learning process. In this case, the network needs to generate a new neuron as its parent's attribute is inherited. Case 2 If a neural network has enough neurons to infer, and even if the input weight vector of each neuron will converge to a certain value, then we shall be able to turn out unnecessary neurons from the network in the calculation. In this case, because it is necessary to delete a redundant neuron to the calculation, the neuron is annihilated without affecting the performance of the network. By observing such behaviors, we can generate or annihilate the specified neuron respectively to achieve an overall good system. In the proposed method, we described a procedure to derive the neuron generation/annihilation automatically and applied the procedure to learning system. Next, we apply such procedure to the learning system in which the experimental data related to hepatobiliary disorders is used. We use a real medical database containing the results of ten biochemical terms test for four hepatobiliary disorders. We have 536 case data, including some errors. After the learning, by using 179 data chosen randomly from database, the proposed system converged to a certain small value and this constructed network has the optimal structure for these teaching data. In addition, we get that the fuzzy rules have some meanings related to the degree of the input weight vector, and the fuzzy rules for hepatobiliary disorders are extracted from the learned network with respect to the degree of input weight vector. Moreover, to verify the validity of the diagnosis of the proposed method, the feed-forward calculation was implemented using extracted fuzzy rules for all databases. As a result, the proposed system correctly diagnosed more than 70%.