A Research On Robust Estimation Of Extreme Learning Machine

Extreme learning machine (ELM) is a type of single layer feedforward neural networks(SLFNs). ELM has some advantages that the learning speed is faster and the structure issimpler than traditional neural networks. But after deeply studied in this algorithm, we foundthat the robustness of ELM is poor, and it is easy to be affected by outliers. In this paper, anovel approach, called RBELM (Robust ELM), which adopted iterative reweighted leastsquares(IRLS) method on ELM, is proposed to enhance the robustness of ELM.The main contents are outlined as following:(1) We introduced neural networks’ development history and review its basic content andlearning method. The principium and characteristics of ELM are deeply expatiated andinvestigated, and its recent researches in the world have been summarized. In addition, wecompared the advantages and shortages of ELM with traditional neural network.(2) In order to enhance the robustness capacity of ELM and decrease the interference ofgross error, a robust estimation theory is expatiated, especially for the M-estimator.Combining with the two theories, we proposed a novel approach, RBELM, which adopted theweighted least square (WLS) method to calculate the output weights of network, in this way,each sample weight can be determined by the size of residuals, and the robustness of ELM isimproved.(3) According to some experiments, the simulation results showed that RBELM provedits robust capacity is superior to ELM in the “SinC” approximation experiment, regressionproblem and classification problem (UCI datasets), and it still keeps the advantage of ELM,such as the simple structure, fast training speed, and good generalization performance, etc.(4) Rotary kiln is a classical multivariate, nonlinear, strong coupling and interferencecontrolled model. According to analysis the process characteristics of rotary kiln, wepreprocess the thermal data of the rotary kiln, and applied RBELM to predict the rotary kiln’scoal feeding trend.The simulation results showed that the performance of RBELM is better than ELM andBP algorithm.