Abstarct: In this applicable thesis, we study the classifying of five mental tasks that have been measured by EEG. Time components of EEG signals have high volume, and also they might have some artifacts that haven’t been recorded from the brain, such as blxinking, or eye movements. So reducing redundancy data and extracting useful data is necessary. In proposed method PCA and CSP have been used for removing redundancy data. But there are some artifacts that don’t remove by these methods. For this, first segment EEG signals into several windows, then apply PCA or CSP on each window distinguishably. Then remove features of redundant windows. But the problem is that how recognize windows that have artifacts. It is difficult, because of artifacts aren’t clear in EEG signals. But we can probe for solutions related to selecting best windows, by Genetic Algorithm (GA), Particle Swarm Optimization (PSO), and forward feature selection methods. But the problem is that all features of removed windows aren’t redundant, but they are removed. Because in static partitioning, the ridge of artifacts isn’t clear. So we should use proposed method for different sizes of windows in multi steps, and finally select best dimensions. We applied proposed method on one dataset that contains five mental tasks. We consider 70% of data for training and 30% for testing system, and we acquire 100% success rate, for 50 windows.