Multi-aggregation process neural networks and learning algorithm

Aimed at the information process problem that the system inputs are multivariate process functions and multi-dimension process signals, this paper proposes a kind of the multi-aggregation process neuron and the multi-aggregation process neural networks model. The inputs and connection weights of multi-aggregation process neuron all can be multivariate process functions, and its aggregation operations include space weight congregation to many input functions and the cumulation of multi-dimension process effect, can simultaneity reflect many multivariate process input signals that effect together in multi-dimension space and the cumulation result of process effect. Multi-aggregation process neural networks are composed of multi-aggregation process neurons and other type neurons according to certain structure connection, in the light of whether the outputs are multivariate process functions or not, the general model of feedback multi-aggregation process neural networks and multi-aggregation process neural networks model which inputs and outputs are all process functions are founded, have the direct mapping and modelling ability to the input/output connection of multivariate process signals. A kind of learning algorithm based on the grads descending with numerical computation integrating which bases on multivariate function base expanding is proposed in this paper, and the simulation experiments results show the adaptability of models and algorithms to the multivariate process signal classification and the simulation problems of multi-dimension dynamic process.EI0148-563