Multilayer-perceptron training method based on bee colony algorithm with learning factor

Abstract

The invention discloses a multilayer-perceptron training method based on a bee colony algorithm with a learning factor. In the method, an idea of asynchronous adjustment of the learning factor is introduced to update a nectar source position of following bees so that the self-adaptive optimization capability of a bee colony is strengthened and single extreme values and a global extreme value of the colony are used to improve a search mode in the manual bee colony algorithm so that search efficiency and performance are optimized so as to achieve an optimal optimization effect of an initial weight and threshold of a multilayer perceptron and improve the predication precision of the multilayer perceptron.

Description

technical field [0001] The invention relates to the fields of neural calculation and intelligent optimization, in particular to a multi-layer perceptron training method based on bee colony algorithm with learning factors. Background technique [0002] In the past two decades, artificial neural networks, especially multi-layer perceptrons trained with BP algorithm, have been widely used in various fields. Theory has proved that by choosing appropriate weights and activation functions, multilayer perceptrons trained with nonlinear functions can approximate and accurately generalize almost any continuous function without considering the prior assumptions of the data distribution. It can simulate highly nonlinear functions and can be trained to generalize accurately on unused new data. Like many neural network models, multilayer perceptrons must be trained to gain fit, prediction, and more. As far as training is concerned, the BP algorithm based on gradient descent is the most...

AI Technical Summary

Problems solved by technology

As far as training is concerned, the BP algorithm based on gradient descent is the most commonly used algorithm for training multilayer perceptrons, but it often encounters some problems when using it, such as high sensitivity of initial weights during training, slow convergence rate, and easy to fall into local best

Therefore, some new training techniques have been gradually proposed, such as the particle swarm optimization neural network method, in an attempt to improve these shortcomings, but due to the high complexity, high-dimensional, and multi-modal optimization problems of the neural network training process, these techniques are far from perfect. satisfactory

Images