Objective To improve the classification performance of the surface electromyography (sEMG)-based prosthesis and reduce the dimensions of features extracted from the sEMG signals，a modified ant colony optimization(ACO)was employed to select the best feature subset. Methods The relationship between features and target classes was calculated as the heuristic function and the best feature subset was selected by ACO，and the trained artificial nerve net was utilized to verify the classification performance. Results Ten healthy subjects participated in the experiment on classification of hand and wrist motion using sEMG signals. Compared to the principle component analysis(PCA)-based feature subsets，the ACO-reduced feature subsets not only improved the classification accuracy but greatly reduced the number of features in the original feature set，which subsequently simplified the structure of the classifier and reduced the computational cost. Conclusions The proposed method exhibits a great potential in the real-time applications，such as sEMG-based prosthesis control.