Adaptive Network Selection Method Ultimate Learning Machine

As an emergent technology, Extreme Learning Machine (ELM) has been developed as a linear learning system with fast learning speed by randomly selecting the hidden variables of a nonlinear learning system. Different from the conventional learning methods, the hidden-node parameters do not need to be adjusted iteratively in ELM. Therefore, ELM can significantly reduce the complexity and hence increase the efficiency. It has been shown that ELM performs well in various regression and classification applications. However, how to determine the optimal network architecture as well as keep the good generalization performance is a key factor of guaranteeing the efficiency of ELM in applications. This thesis proposes the Bisection Extreme Learning Machine with Adaptive Growth of Hidden Nodes (BAG-ELM) and the Improved Dynamic Extreme Learning Machine based on the LU Factorization (LUD-ELM). The main works of this thesis include:Chapter1systematically introduces some background knowledge of the ar-tificial neural networks (ANNs), and the basic idea of ELM as well as its devel-opment.Chapter2first gives a brief overview of ELM algorithm. Then three different classes of network architecture selection approaches, including the constructive algorithms, the destructive algorithms and the adaptively growing algorithms are reviewed systematically.Chapter3proposes the bisection extreme learning machine with adaptive growth of hidden nodes (BAG-ELM) based on the bisection method. The per-formance of the proposed algorithm is verified in six real-world benchmark data sets. Chapter4proposes the improved dynamic extreme learning machine based on the LU Factorization (LUD-ELM). The performance of the proposed algo-rithm is verified in six real-world benchmark data sets.