Parallel Auto-clustering Algorithm Based On CUDA And GEP

Clustering algorithm has been got extensive attention in academia and industry. It has been applied in pattern recognition, data mining, and other important areas; and has achieved fruitful research results. But clustering algorithm has weaknesses like sensitive to the initial clustering center, easy to fall into local optimum and clustering process too slow, these weaknesses influence the precision and convergence speed of clustering algorithm, and limit it’s further used. For this purpose, evolutionary algorithm has been used into clustering algorithm. By means of self-organization and global search ability of evolutionary algorithm, sensitive to the initial clustering center and easy to fall into local optimum problems about traditional clustering algorithm are solved, and good results are obtained. With the increasing of data scale, convergence speed of evolutionary clustering algorithm has become the main factors influencing the performance of this algorithm. According to this problem, in order to improve the convergence speed of the algorithm, using the clustering techniques as the main approach to parallel evolutionary clustering algorithm up to now, and also achieved good effect. However, although clustering techniques provide good parallel performance, it consumes a lot of resources. It can not be widely used because of its exorbitant cost. As a kind of multi-core parallel computing architecture, compute unified device architecture (CUDA) has good parallel performance, and just need a single computer equipped with a CUDA supported GPU. Based on this, the CUDA architecture is applied in evolutionary clustering algorithm in order to improve its performance. At the same time, gene expression programming (GEP) algorithm which is a member of evolution algorithm, whose ability of self organization and global optimization are good to improve traditional clustering algorithm. On the basis of traditional parallel GEP model we designed three parallel GEP model based on CUDA, then used these models to design and implement automatic clustering algorithm. Experimental results show that the proposed automatic clustering algorithm has better performance than the traditional automatic clustering algorithm.