K-means clustering algorithm is one of the main algorithms applying in machine learning and pattern recognition. However, as the center of clusters are selected randomly and also due to the dependence of clustering result on the initial centers of clusters we may trap into local optima centers. In this paper a new genetic algorithm approach based on k-means algorithm is suggested in which the centers of clusters are selected better and in an appropriate manner. In order to increase the efficiency of this algorithm, in each stage, the layout of cluster centers which are in the form of chromosomes are changed with respect to the best chromosome. By estimation of results of the proposed approach on a standard data set and also comparison of this algorithm with other related algorithms we can show that our approach is more efficient than k-means algorithm and other algorithms which have been selected in this paper for comparison purposes.