Evolutionary extreme learning machine based on dynamic Adaboost ensemble

Abstract

Boosting ensemble algorithm exhibits two fatal limitations: one is that it gives in advance the upper bound of weighted error on weak learning algorithm; the other one is that it is overdependent on data and weak learning machine, and it is too sensitive to data noising. Aimed at limitation of Boosting ensemble application in extreme learning machine, this paper proposes a new algorithm: evolutionary extreme learning machine based on dynamic Adaboost ensemble，which regards the evolutionary extreme learning machine as weak learning machine, dynamic Adaboost ensemble algorithm is used to integrate the outputs of weak learning machines, and makes use of fuzzy activation function as activation function of evolutionary extreme learning machine because of low computational burden and easy implementation in hardware. Proposed algorithm has been successfully applied to problem of function approximation and classification application. Experimental results show that the algorithm increases the training speed greatly when dealing with large dataset and has better generalization performance compared to extreme learning machine, evolutionary extreme learning machine and Boosting ensemble extreme learning machine with quasi-Newton algorithms.