A self-adjusting and probabilistic decision-making classifier based on the constructive covering algorithm in neural networks

It usually needs complicated nonlinear operations to get the characteristics from the raw information inputted, and it is very difficult to find this kind of algorithm directly. The geometrical meaning of the multilayer perceptron´s neuron model indicates that classifying samples according to the requirements by constructing neural networks is equal to finding a collection of domains with which vectors of the preset sample sets are partitioned. But in some applications, such as time series forecasting including stock share forecasting, due to their preset sample sets may contain some exceptions and erroneous results, it is desired to introduce some self-adjusting and probabilistic decision-making mechanism to enhance the accuracy of classification. At the same time the mechanism can reduce the size of neural networks and speed up the recognition process. We discuss a self-adjusting and probabilistic decision-making mechanism for the covering algorithm. Based on the method, we developed a self-adjusting and probabilistic decision-making classifier and applied the software package to forecast the share index of Shanghai´s stock market.