Researches On A Classifier Based On Extreme Learning Machine And Its Application

Discovering knowledge from huge data is the basis of complex data analysis and decision-making system establishment. Pattern classification is one of important research issues in knowledge discovery. In recent decades, artificial intelligence (AI) methods have been widely applied into many real-world classification problems, such as medical diagnosis, mechanical diagnosis, speech recognition, face recognition and etc.Extreme learning machine (ELM) is a new-generated AI method, which was put forward by Huang in 2006. As one of single hidden layer feed-forward neural networks, ELM's learning algorithm can achieve the output layer weights using Moore-Penrose generalized inverse once generating randomly the input layer weights. Comparing with the traditional neural network learning algorithm, ELM learning algorithm is easier and faster to execute since the achievement of the unique analytic optimum only depends on setting the number of hidden layer nodes, while the input layer weights and the bias of hidden layer nodes do not require to be adjusted during the running course of algorithm. Due to its robustness, ELM has gained a lot of concerns from AI community in recent years.Based on the mechanism of system engineering, this thesis utilizes the relevant knowledge and methods from the fields of AI, information science and statistics to study the algorithm principle and applications of ELM. The main contents of this thesis can be organized as follows.(1) Reviews on the relevant research works. The development periods of AI and several classical Al-based classification methods are firstly introduced in brief and then the research status of ELM are described in detail.(2) Studies on the algorithm principle of ELM. Four important parameters, that is, the number of training samples, the number of hidden layer nodes, the values of input layer weights and the setting ways of hidden layer nodes'bias, are examine and test their influence degrees upon the performance of ELM classifier through the simulation experiments on a Benchmark classification problem. In addition, the normalization method of training samples is also to test its validity due to the requirement in the practical applications.(3) Studies on the practical applications of ELM. An ELM classifier is designed for a practical application problem on the Fall-Down test and its effect and accuracy degree on an example, which is generated from a set of actual data, are investigated with the analysis of parameters' sensitivities.(4) The research works in this thesis are summarized with discussions on the future work.