Evolving Neo-Fuzzy System for Distorted Data Online Processing

The problem of evolving system of computational intelligence synthesis for online processing of stochastic nonstationary sequences distorted by missing values in the data is considered when these missing values can be present both in the input signals and in the reference one. The evolving neo-fuzzy system, that in the process of operation restores the missing values in data, adjusting its parameters (synaptic weights) and architecture is proposed. The elements of the system are nonlinear synapses of the neo-fuzzy neuron, where the traditional membership functions are replaced by orthogonal polynomials. Using of orthogonal polynomials allows to rebuild the architecture of the system without changing the already formed membership-activation functions and the corresponding synaptic weights. For the tuning of system parameters we use either the recurrent least squares method or the multistep gradient learning algorithm with the sliding window both. The neo-fuzzy system under consideration is designed to solve the Data Stream Mining tasks including prediction, adaptive identification and control of nonlinear stochastic and chaotic objects.