Research On Regression Algorithm Of Extreme Learning Machine Based On Robust Improvement

The standard extreme learning machine is a neural network with strong fitting ability,simple model structure,good generalization performance,and fast learning speed.It has been widely studied and applied by many domestic and foreign scholars in recent years.However the standard ELM algorithm also has some shortcomings.In the process of training the ELM model,when there is multicollinearity between the column and the column of the hidden layer design matrix,the hidden layer design matrix will present a morbid condition,which leads to the deterioration of the model generalization performance and robustness.In response to this situation,many scholars have proposed some improved algorithms,such as the ridge regression extreme learning machine,and an extreme learning machine based on principal component estimation.Based on the ridge regression ELM algorithm,this paper analyzes and studies the network structure and algorithm of ELM model,and proposes an extreme learning machine regression algorithm(CV-ELM)based on conditional index and variance decomposition ratio.This article mainly carried out research work from the following directions:(1)It briefly introduces the background,practical significance and current research status of the Extreme Learning Machine,and introduces and analysis of the standard extreme learning machine and various extreme learning machine theories based on robustness improvement,such as ridge regression extreme learning machine(RELM)and extreme learning machine based on principal component estimation(PC-ELM).The dsadvantages and deficiencies of the standard extreme learning and the ridge regression extreme learning machine are mainly analyzed.(2)Through the analysis of the standard ELM and the algorithm flow of the extreme learning machine based on the robustness improvement,this paper presents a extreme learning machine regression algorithm based on the condition index and the variance decomposition ratio.The algorithm first uses conditional index and variance decomposition to preprocess the hidden layer design matrix of the model,separates the interference items in the design matrix,then uses the ridge parameter to weigh the interference term,and finally calculates the output weight by the least square estimation method.The experimental results show that the method is targeted to deal with interfernece items and avoids the impact of the ridge regression method on the non-interference items in the design matrix,and the training accuracy of CV-ELM has been improved to a certain degree under the condition that CV-ELM is equivalent to the robustness of ridge regression extreme learning machine.In order to verify the robustness of CV-ELM algorithm,we have done many comparative experiments on different scale datasets.(3)The extensive application of the extreme learning machine algorithm is partly due to the random input weight and the hidden layer bias,and the input and hidden layer bias of the CV-ELM algorithm are also randomly generated,which will lead to poor robustness of CV-ELM algorithm.In order to overcome this shortcoming of CV-ELM,this paper introduces ensemble learning method,and proposes a CV-ELM regression algorithm based on ensemble learning(ECV-ELM).This method uses the ensemble learning method to generate several CV-ELMLlearners,and then interates the optimal set of sub-learners through the averaging method to obtain a good network model.The experimental results show that the ECV-ELM algorithm can make some CV-ELM learners complement each other,so that the generalization performance and robustness of the model have been improved,and the number of ensemble sub-learners is not as much as possible.