Fractional-order neural network modeling method based on smooth Group Lasso penalty term

Abstract

The invention discloses a fractional-order neural network modeling method based on a smooth Group Lasso penalty term, and the method comprises the following steps: selecting a neural network model, and determining an error function; Updating a training network weight through employing a fractional-order steepest descent method, i.e., updating the weight value in a fractional-order negative gradient of the error function relative to the weight value; obtaining the network parameters of a neural network model according to the fractional-order steepest descent method at step 2; and calculating the precision of the neural network model through employing a testing sample. The beneficial effects of the invention are that the method carries out the training of the network weight value through employing the fractional-order steepest descent method; compared with an integer-order derivation gradient algorithm, the method is higher in precision because the fractional-order model description is more accurate than integer-order model description.

Description

technical field [0001] The invention relates to the field of neural network model modeling, in particular to a fractional order neural network modeling method based on smooth Group Lasso penalty items. Background technique [0002] At present, the error backpropagation (Back Propagation abbreviated as BP) neural network is a multi-layer forward network. The inter-layer neurons realize the full connection of weights, and there is no connection in the layer. For the learning of its weights, the gradient is the most used. The descent method is to calculate the partial derivative of the objective function for each weight according to the error between the ideal output and the actual output, and modify the weight in the opposite direction of the partial derivative to achieve the purpose of continuously reducing the output error. In addition, the BP algorithm often uses methods such as conjugate gradient (conjugate gradient) method and Gauss-Newton (Gauss-Newton) method to realize...

AI Technical Summary

Problems solved by technology

[0006] Although BP neural network is widely used, it has three essential shortcomings: slow convergence speed, poor network fault tolerance, and easy to fall into local minimum without obtaining the global optimal solution.

[0007] Literature [2] proposed a neural network model based on the fractional steepest descent method. This algorithm helps to improve the convergence accuracy of the network, but the number of neurons and connections is redundant, and the structure of the neural network is complex and does not have sparsity.

Images