Fuzzy backpropagation networks using vector valued activation function with applications to data fusion.

摘要

This research is to construct architectures, develop learning methods and design the related algorithms for fuzzy backpropagation (FBP) multilayered neural networks (NN) that have n x m input space and use the vector valued activation functions at various layers. They are the extension of the standard fuzzy backpropagation NN taking not only binary, but also any values from the interval [0,1].;Supervised training methods that was used in the new algorithms involves the following stages: fuzzification of crisp data, the feedforward of the fuzzy set input training pattern, the FBP associated error, the adjustment of the weights, and defuzzification of the fuzzy output.;The main motivation for the establishment of these novel neural networks is to support real life problems in the area of data fusion. Applications include a new multidimensional problem (Body Mass Index (BMI) problem), a known problem (OR logic function) and a previously unsolved problem (XOR logic function) due to nonlinearity. They have proven the higher performance, faster convergence and better accuracy of these new algorithms in comparison with the existing ones.