The algorithm was implemented using precomputed features on data set V of BCI competition III. Researchers: Changshui Zhang, E-mail: zcs@mail.tsinghua.edu.cn Shiliang Sun (corresponding person), E-mail: sunsl02@mails.tsinghua.edu.cn Feiping Nie, E-mail: nfp03@mails.tsinghua.edu.cn Affiliations: Department of Automation, Tsinghua University, Beijing 100084, China. Description: we select psd features corresponding to Mu rhythm from the given 96 features, and retain the training sessions with good generalization as training set. The trained SVM classifier with RBF kernel function is applied to classify test sessions. The parameters of SVM classifier are selected through 10 fold cross-validation on the corresponding training set. As a result, for subject 1, the training set includes session 1 and 2. Since classifier trained on session 1 or session 2 provides good classification performance on other two training sessions. The features used are psd of 8, 10, and 12 Hz of all 8 channels. As these features reflect Mu rhythm activity from our priror knowledge and experience and would be discriminative for classification task. Similarly, for subject 2, the training set includes session 2 and session 3. The features used are psd of 8, 10, and 12 Hz of all 8 channels. For subject 3, the training set includes session 1 and session 2. The features used are psd of 10, 12, and 14 Hz of all 8 channels. The final classification result is recorded in file OutputV.mat. OutputV.mat (using Matlab 6.5) contains the responses for every 8 input vector in test sessions (give an output every 0.5 s). There are three variables in the output file, which are 'Subject1_TestY', 'Subject2_TestY', and 'Subject3_TestY' corresponding to the estimated labels for the test sessions of subjects 1, 2, and 3 respectively.