Parallel Implementation Of Membrane Computing For Data Clustering On GPU

Clustering analysis is an unsupervised learning processure.The traditional clustering algorithm of K-Means and FCM can be as a representative,however the two algorithms are sensitive to initial seeds and noise sample points.All kinds of ills lead to a class of algorithms based on membrane computing have been proposed and used,which can effectively solve these problems and shortcomings.Membrane clustering algorithm is a novel membrane computing-inspired clus tering algorithm,whose key component is a P system.The P system is a distributed and parallel computing model.However,the current computers are with serial architecture,so the algorithms realized by general CPU is the serial analog,which can't reflect the parallel computing advantage of P systems.The birth of heterogeneous GPU supercomputing brings a breakthrough for distributed parallel computing model,especially the CUDA on NVIDIA GPU(Unified Device Programming Model)is striking.It uses a double parallelism highly parallel model and the characteristics of the programming model fit to P system(parallel in cells and parallel in different objects' evolution within cells).At the same time,the multi-level storage system of GPU hardware just meets the needs of the mechanisms that P systems share/communicate the objects between the cells.In this paper,using the tissue-like P systems as the basic computing framework and incorporating the particle swarm mechanism/improved particle swarm mechanism as the evolution rules of object,two kinds of membrane clustering algorithms,Hard-MC and Fuzzy-MC algorithms,are proposed for the hard clustering and fuzzy clustering problems.In order to take full advantages of P systems,two kinds parallel implementation framework of P systems is proposed according to the architecture of GPU.Based on the implementation framework,the parallel versions of the two membrane clustering algorithms are discussed and implemented.Finally,the comparison results on several benchmark data sets have validated the availability of the parallel algorithms and the advantages of parallel computation.