Neural network increment-type feedforward algorithm based on sample increment driving

Abstract

The invention provides a neural network increment-type feedforward algorithm based on sample increment driving. According to characteristics of input samples, input weights and hidden layer thresholds of a neural network are obtained by a least square method, and initialization of model parameters is completed; and the input weights and hidden layer thresholds of the model are updated according to increments between new samples and old samples, function relations between the model parameters and input samples are established, and functions of sample self-adaptability and online feedforward adjustment are realized. The algorithm provided by the invention has the advantages that the prediction precision is high, the generalization capability is high, and online feedforward adjustment is available.

Description

technical field [0001] The invention relates to the field of online learning neural network, in particular to an algorithm for online learning single hidden layer feed-forward neural network. Background technique [0002] Since the 1980s, neural networks have derived a variety of model structures and learning rules, such as the earliest proposed BP neural network based on error backpropagation learning, radial basis neural network, discrete and continuous Hopfield neural network , SOM neural network, etc. The various neural networks described above can be used for both classification and regression problems, and all have achieved good model results. The extreme learning machine algorithm proposed in recent years has become a research hotspot. It belongs to a single hidden layer feedforward neural network. The single hidden layer feedforward neural network has good nonlinear identification ability, and has simple structure, few adjustable parameters, and learning With the a...

AI Technical Summary

Problems solved by technology

[0003] At present, all practical engineering application system modeling is based on time series data, and the input weights and hidden layer thresholds of most neural networks are randomly initialized, and the generated model parameters are consistent with the input sample data of the actual system. There is no inherent regularity, which leads to poor predictive effect and stability of the neural network

Images