Research On Regression And Classification Methods Based On Multiple Parallel Extreme Learning Machine

Compared with the traditional neural network,Extreme Learning Machine(ELM)has excellent performance in speed and computation.It realizes the complex nonlinear mapping directly from the input layer,and transforms the iterative solution process into the solution process of linear equations by the method of randomized parameters,which makes it obtain faster solution speed.At the same time,it avoids the problem of easily falling into the local optimal solution.However,there are also some shortcomings in ELM,such as the need for more hidden layer neurons and sometimes the problem of sacrificing meaningful features of data to improve training speed.In order to improve the accuracy and stability of regression and classification,we studied and analyzed the problems of extreme learning machine in different regression and classification data sets,and achieved the following results:First,aiming at the neglect of traditional neural network on inhibitory neurons,a network model of multi-parallel Extreme Learning Machine based on Excitatory and Inhibitory neurons(MEI-ELM)is proposed.Considering the advantages of parallel,based on Fast Learning Network with Parallel Layer Perceptron(PLP-FLN),we propose a multi-parallel Extreme Learning Machine.On this basis,utilizing the characteristics of biological neurons to transmit information,we introduce neurons with inhibition and excitation in the network.The MEI-ELM is compared with the other three networks on nine classic benchmark regression problems to test the effectiveness of the network.The results show that the prediction accuracy of MEI-ELM is at least 2 times higher than other networks.Second,in order to make full use of the effective information in the network errors,based on the idea of error compensation,an improved MEI-ELM——Multi-parallel Extreme Learning Machine based on Error Compensation,MEC-ELM,is proposed.Similarly,through 9 regression problems,MEC-ELM shows better generalization ability than other algorithms in most data setsIn addition,in order to verify the effectiveness of the improved MEC-ELM in the classification problem,we also applied it to 6 classicclassification problems,which improved the classification accuracy by an average of 3.5%compared with the other 5 networks,and had a higher classification accuracy.Finally,aiming at the problem of data online training,based on the MEC-ELM model,an Online sequential Multi-parallel Extreme Learning Machine(OMEC-ELM)is proposed.The simulation results show that OMEC-ELM can achieve good prediction and classification accuracy under the condition of sigmoid function.