A Study Of Ensemble Extreme Learning Machine Based On The Selection And Optimization Of Member Classifier

Extreme learning machine has been widely used because of its good performance. The ensemble ELM has the better characteristics of convergence and generalization performance than the single ELM. However, random initialization input weights and bias of ELM leads to low stability of the algorithm, which directly affects the diversity and accuracy between classifiers in the ensemble ELM. Therefore, in the paper, K-means is applied to cluster the ELMs efficiently while tournament-selection is used to choose the optimal base ELMs. Then, ELM cluster further divided into two groups, ARPSO is applied to optimize the two populations depending on the different convergence respectively. The main work of this paper is as follows:(1) A method of selecting and clustering base ELM based on K-means and genetic algorithm— KGA-DOEELM is proposed. In the proposed algorithm, based on the input and output weights of ELM, K-means algorithm is applied to cluster the base ELMs rapidly in the first phase, which greatly increases the diversity between the different classes of samples. Then, tournament-selection is used to choose better members in different clusters, which further improves the classification accuracy of ensemble system. Experimental results confirm the better diversity and accuracy of the proposed algorithm in ensemble system.(2) On the basis of KGA-DOEELM, the ensemble of ELM based on the multi-subpopulation is proposed—M-KGA-DOEELM. The proposed algorithm is aim to supplement the missing members after initial selection. Firstly, after clustering the optimal ELM sets by K-means method, tournament-selection is applied to select the base ELM into the best group and general group in turn by fitness value. Two groups still follow the steps ARPSO optimization iteration. However, in suboptimal group, it adjusts inertia weight in the process of using ARPSO iteration according to the evaluation criterion of convergence. The ensemble of ELM based on K-means and genetic algorithm of multi-subpopulation can ensure the optimal convergence and accuracy. Apply the proposed ensemble ELM to data classification, experimental results show that the proposed ensemble of extreme learning machine in the paper can achieve the high stable performance and accuracy effectively.