Evolutionary Neural Fuzzy Systems for iltering
F. Russo · 1998
A new class of neural fuzzy filters for removing noise from 2-0 measurement data is presented. The proposed approach combines the advantages of fuzzy and neural paradigms. The network structure is, in fact, specifically designed to exploit the effectiveness of fuzzy reasoning in removing noise without destroying the useful information embedded in the input data. Easy design of new filters is thus obtained because the neuro-fuzzy approach is capable of automatic acquisition of knowledge for a given network structure. The learning method based on the Genetic Algorithms performs an effective training of the network yielding satisfactory results after a few generations. Experimental results show that the proposed approach is very effective also in presence of data highly corrupted by noise. The neural fuzzy system is able to largely outperform other methods in the literature including state-of-the-art techniques. classification and control purposes (I 5-1 81, the proposed system adopts a specifically developed network structure. Fuzzification of input data is performed by resorting to 2-D fuzzy sets. This information is then processed by means of trainable fuzzy aggregators. The learning method is based on the Genetic Algorithms (GAS). GAS, which search for the optimal solution of a problem by applying the mechanisms of natural selection and natural genetics, have demonstrated in fact to be very effective for the training of neuro-fuzzy systems 119-221. Many computer simulations have been carried out to assess the performance of the proposed approach. Extensive validation of the method is presented using different collections of 2-D data corrupted by noise. Experimental results show that the proposed neural fuzzy system is very effective in removing noise and performs significantly better than state-of-the-art filters in the literature. This paper is organized as follows. Sect.11 describes the structure of the neural fuzzy system, Sect.111 focuses on the encoding of the network parameters and the genetic learning, Sect.lV shows the experimental results and, finally, Sect.V reports conclusions.