Anytime Extension of Generalized Fuzzy Neural Network Models with Non-singleton Consequents

Nowadays practical solutions of engineering problems involve model-integrated computing. Model based approaches offer a challenging way to integrate a priori knowledge into the procedure. Recently, Artificial Neural Networks (ANNs) has become popular because they are able to learn complex mappings from the input/output data and are relatively easy to implement in any application. Generalized Neural Network (GNN) based models may have an exceptional role at many fields, where the problem to be solved is complex, highly nonlinear or when only partial, uncertain and/or inaccurate data is available, however their high complexity, and in certain cases unknown accuracy, can limit the applicability, especially in time critical situations. Combining GNNs with anytime techniques may offer a solution to the complexity problem, but only in cases when the error of the reduced models can be estimated. In this paper, author gives error bounds for a new class of GNNs where the nonlinear transfer functions are approximated by product-sum-gravity fuzzy systems with non-singleton consequents (NGFNN) thus extending the range of possible anytime soft computing models. The model complexity can flexibly be reduced to cope with the temporal, possibly dynamically changing resource, time, and data availability. It is shown that the accuracy of the reduced models can be upper-bounded and thus, the error is always known, and further, monotonously decreases parallel with the increase of the complexity of the used model. These features make NGFNNs suitable for anytime use.