Abstract
Recently, researches on brain wave based controls of robots have been reported. They are achieved by thought recognitions using brain waves. The authors have proposed a method that transforms brain waves to frequency series and apply a discriminant analysis to recognize one's thought. The conventional method uses a variable selection by a cross-validation to prevent an over-training. The number of candidates of variables is, however, a product of that of measuring locations times that of frequency series. As a result, it is impossible to calculate all the combinations of them within a real-time. This paper proposes a new variable selection method based on p-values in the t-test. It does not take much time to calculate p-values. In addition, it is expected that the proposed method will improve the recognition accuracy because it can select variables from the candidates that are lost in the conventional method. It is also expected that it makes possible a real-time training and therefore it enables us to feed back information such as fatigue and accuracy to subjects during measurements.
