Abstarct: Development of Brain–computer interface systems have attracted many attentions from scientists in the last two decades. A brain-computer interface (BCI) is a system which receives signals such as electroencephalogram (EEG) from brain and converts them into proper commands to control peripherals such as an implant arm, a wheelchair, a robot or other accessory devices. In brain-computer interface baxsed on motor imagery, a person is asked to imagine some specific movement of his body, without actually moving any part of his body. This imaginary movement will lead to producing special signals in the brain which effects on EEG signals. This thesis is concerned with classifying the four classes of motor imagery using EEG data. The considered classes are motor imaginary for right hand, left hand, foot and tongue. The data set used for evaluating the proposed system in this thesis is Dataset 2a of BCI Competition IV. This competition was held in 2008. In our study, we suggested two methods for BCI classification. One of well known methods for discriminating imaginary movements is Common Spatial Pattern (CSP) method. In this method, selecting a proper frequency filter along with a spatial filter is an important challenge. Unfortunately the proper filter band differs from a person to another. In one of the recent methods called OSSFN (Optimal Spatio-spectral Filtering Network), the optimization of spatial filters jointly with a bandpass filter is proposed baxsed on maximization the mutual information between the feature vectors and the corresponding class labels. In the first proposed algorithm of this thesis, we extended OSSFN method from a two-class issue to a four-class solution. Synchronized optimization of frequency-spatial filters will increase the accuracy of classification comparing to CSP method. In the second part of this thesis, instead of considering mutual information between the feature vectors and the corresponding class labels, KAPA criterion is used for optimization. KAPA criterion is measure of classification accuracy. For optimization, we used genetic algorithm rather than gradient method which is highly complicated and likely to be trapped in local optimum points. The result of experiments indicates the superiority of recommended method to OSSFN. The proposed algorithm is called GAOSSFN.