Recurrent fuzzy neural computation: Modeling, learning and application

A novel recurrent neurofuzzy network is developed in this paper. The network model is composed by two structures: a fuzzy system and a neural network. The fuzzy system contains fuzzy neurons modeled using t-norms and s-norms. The neural network is composed by nonlinear elements placed in series with the fuzzy system. The network model implicitly encodes a fuzzy rule-based system and its recurrent multilayered structure performs fuzzy inference. The topology induces a clear relationship between the network structure and the associated fuzzy rule-based system. Network learning involves three main steps. The first step uses a modified vector quantization approach to granulate the input universes. The next step assembles the network connections and their initial, randomly chosen weights. The third step uses two main paradigms to update the network weights: gradient descent and gradient projection method. The recurrent fuzzy neural network is particularly suitable to model nonlinear dynamic systems and to learn sequences. Computational experiment with a classic prediction problem benchmark shows that the fuzzy neural model outperforms a finite impulse response neural network.