Study On The Structure Optimization Of Extrelem Learning Machine And Its Application

The tradition BP gradient learning method causes some problems, such as long training time, over-fitting and easily falling into local optimal. The extreme learning machine (ELM) has a faster learning speed. The algorithm is simple and can be easily implement. It has a better generalization performance and the capability of avoiding over-fitting. It has widely used in regression and classification applications. But the extreme learning machine has a more complex network structure and a slower response speed to the unknown data than BP neural network. This paper is focus on structure optimization of ELM and the main work is as follows:1. Thinking that the reason is randomly choosing input weights and biases and cannot make sure it effective after researching on complex network structure of ELM. So an incremental extreme learning machine based on recursive least square method (RLSI_ELM) is proposed which introduces recursive least square (RLS) to calculate the input weights and biases of extreme learning machine. The experiments results show that RLSIELM can gain a more compact network structure in a shorter training time than EI_ELM and EM_ELM in the same classification accuracy.2. Considering the strong optimization ability of quantum particle swarm optimization, an extreme learning machine based on quantum particle swarm optimization (QPSO_ELM) is proposed which introduces quantum particle swarm optimization to gain the input weights and biases of extreme learning machine instead of randomly choosing. The experiment results show that QPSO_ELM can obtain more efficient input weights and biases and simpler network structure than ELM.3. QPSO ELM is used in handwritten numeral recognition to construct the suitable numeral classifier. The algorithm is tested for classification ability on USPS handwritten numeral recognition. The experiment results show that QPSO_ELM can have a better Classification accuracy than ELM and BP neural network. And its training time is less than handwritten numeral recognition based on BP neural network.