Abstract

This work describes the neural net as identification process, considering an adaptation that adjusts the parameters in dynamical form using a proportional condition. Neural nets have problems when the weight gains are dynamically modified in a specific environment with stochastic smooth conditions adjusting in some sense in accordance with a reference signal. This output and identification error has distribution functions described as fuzzy logic stochastic actions, building membership functions with mobile inference limits, adapting the neural net weight processes in accordance with inference mechanisms. Specifically, the identification process is based on: a) A black-box scheme, where the internal weights are unknown; b) The adaptive criterion that dynamically adjusts the internal weights using fuzzy logic stochastic properties bounded by a transition function without stability loss conditions. The adaptive filter operation was illustratively presented.

Key words: Estimation, fuzzy systems, neural net, digital filter, identification.