Spike sorting involves clustering spikes according to the similarity of their shapes. Usually the sorting procedure is carried out by extracting appropriate features of neuronal spikes. In this study a new spike sorting procedure based on genetic algorithm is developed which contains two distinct phases. In the first phase a B-spline curve is fitted to each spike waveform and then the optimal features are selected from parameters of fitted B-spline curves. The genetic algorithm is used for searching the optimal parameters of B-spline curve in a way that the curve fitting error is minimized. In the second phase, clustering of spikes based on extracted features is performed by applying genetic algorithm. In this phase the fitness function is defined in a manner that both spatial distances between objects in the feature space and their similarity in the real world are considered. The proposed sorting method is tested on the real neural dataset which firstly are classified by an expert human. The results show that the proposed method based on genetic algorithm framework gives fewer errors of clustering in comparison with some other approaches currently used in the clustering purposes.