Support Vector Machine (SVM) is a new technique for pattern classification, and is used in many applications. Kernel parameters set in the SVM training process, along with feature selection, will significantly impact classification accuracy. The objective of this paper was to simultaneously optimize parameters while finding a subset of features without degrading SVM classification accuracy. A simulated annealing (SA) approach for feature selection and parameters optimization was developed. Several UCI datasets are tested using the SA-based approach and the grid algorithm which is a traditional method of performing parameter searching. The developed SA-based approach was also compared with other approaches proposed by Fung and Mangasarian, and Liao et al. Results showed that the proposed SA-based approach significantly improves the classification accuracy rate and requires fewer input features for the SVM.