Research On Extreme Learning Machines Optimization Methods

Extreme learning machine is presented in recent years as a kind of novel machinelearning algorithm. In the case of input weights and hidden layer offset random set, allthe parameters of the network need not iteration adjustment, but by seeking aminimum error two norm least-squares solution to obtain. Because when the hiddenlayer parameters were randomly set, the output layer weights can be obtained inanalytic form solutions, so it is able to achieve very fast learning. This method has thefeatures of simple principle, fast learning and good adaptability.Although the extreme learning machine is simple and fast, but its input data onthe network is still required when faced with massive data, or when too many outliersin the data will affect the network training effect, this needs to carry on theoptimization of network structure. Meanwhile, the selection of kernel functions forextreme learning machine is also a problem worthy of discussion, and then there isalso the kernel function optimization problem.To solve the above problems,this papermainly studied on the optimization for the extreme learning machine algorithm. Theoptimization focus on the model of network structure and the kernel function of ELM.The main contents of the article generally as follows:1. Combining the attribute reduction algorithm of rough set based on decisiontables, we add a layer of data preprocessing to the original extreme learning machinenetwork structure. This layer is located in the middle of the input layer and the hiddenlayer. The input data samples for attribute reduction, removes redundant properties,reduce the sample dimensions.2. Proposed an algorithm named extreme learning machine optimization basedon wavelet kernel function and use it to train the network. Considering the influenceof Kernel Function of Extreme Learning Machine, and compared with the traditionalspeed Machine Learning Algorithm has higher precision of learning.3. Proposed an algorithm called extreme learning machine based on hybridkernel function. The hybrid kernel combines the advantages of both the local kernelfunction and the global kernel function. The combination hybrid kernel function wasapplied to the extreme learning machine algorithm as a kernel function. And then, weuse this algorithm to train the samples. This is an optimization for the ELM kernelfunction which making the learning performance and the generalization performanceof the algorithm has been improved at the same time.