Abstarct: One of the challenges in pattern recognition is the data attribution selection process. Feature selection plays an important role in solving problems with high-dimensional data and is an important and fundamental step in pre-processing many classification and machine learning problems. The proposed method of this dissertation is a feature selection method baxsed on the gray wolf algorithm. In this algorithm, the Wrapper method is used to select the feature selection. The transfer function is also an important part of the binary gray wolf algorithm, which is necessary for mapping a continuous value to a binary value. In all previous research, only one transfer function was used for the whole algorithm, and all wolves in the whole algorithm dealt with this transfer function. In this dissertation, the algorithm deals with several transfer functions and each wolf will have its own transfer function. The idea stems from the fact that algorithms are evolutionary mexta-innovations and can optimize themselves. Do not depend on the transfer function. In the proposed method, eight transfer functions are used, which are divided into two families, S-shaped and V-shaped. Two approaches are proposed for learning the transfer function. In the proposed method, each wolf is assigned a specific function. This idea can make it possible to select the appropriate transfer function in each iteration of the algorithm according to the selection of the alpha wolf and the movement towards the prey. We used two ideas in this innovation. In the first idea, we add three or two binary bits to the initial population. If two bits are added, four modes of the transfer function are available, and if three bits are added, eight transfer functions are available. In the second idea, 10 or 21 binary bits are added to the initial population. If 10 bits are added, we will have a type of transfer function. 210 coefficient modes are available for the slope of the transfer function. If 21 bits are added, two types of transfer functions are available, of which there will be 210 coefficient modes for the slope of the transfer function. In both ideas, the added standard bits are used to select the transfer function as well as the coefficient affecting the slope of these functions. After each iteration, the alpha wolf position algorithm is updated and the transfer function is selected. With subsequent iterations, the algorithm, while selecting the appropriate transfer function with the appropriate slope, also learns the transfer function. Experimental results on ten UCI datasets show that the selection of the obtained feature subset with high classification accuracy is effective and efficient.