Research On Dynamic Network Structure Selection Method Of Extreme Learning Machine

The architecture of single-hidden-layer feedforward networks (SLFNs) di-rectly affects the generalization performance and computational complexity of a learning algorithm. An SLFN with too few hidden nodes may lead to high training error and poor generalization performance due to underfitting and large statistic bias, while an SLFN with too many hidden nodes may still lead to poor generalization performance due to overfitting and large variance. There-fore, how to find an appropriate and parsimonious network is rule of thumb to obtain good generalization performance. This thesis proposes two novel algo-rithm, EMG-ELM and PMN-ELM, where the network structure are adjusted dynamically. Specifically, adding to hidden nodes according to the arithmetic progression to determine the scope of the optimal number of hidden nodes at the first step. Then, removing those inactive hidden nodes, on the basis of the golden section method and the minimum norm principle respectively, until the optimal number of hidden nodes is found. The main work of this thesis includes:Chapter 1 introduces the background and development of artificial neural network and ELM.Chapter 2 systematically introduces ELM with adjustable architectures, including the constructive algorithms, the destructive algorithms and the adap-tively growing algorithms.Chapter 3 proposes the golden-section extreme learning machine based on error minimum (EMG-ELM). Firstly, adding to hidden nodes according to the arithmetic progression to determine the scope of the optimal number of hidden nodes; Secondly, removing those inactive hidden nodes, on the basis of the golden section method.Chapter 4 proposes a pruned extreme learning machine based on the min-imum norm (PMN-ELM). In this algorithm, the growth stage is the same with EMG-ELM, in the deletion stage using the minimum norm principle find the optimal number of hidden nodes.