The Parameter Optimization of Gaussian Function via the Similarity Comparison within Class and between Classes

Gaussian function is widely used as similarity measurement or kernel function in pattern recognition. The quality of application depends on the parameter selection of Gaussian function. The main method of parameter selection for Gaussian function is cross validation, which is time-consuming for large size of optimization problem. A new heuristic approach is proposed in this paper, which is based on the similarity comparison of training samples within class and between classes. The validity is demonstrated by the experiments on some artificial datasets and benchmark datasets via SVM method.