Parallel Implementation of Matrix Techniques

M. Salmerόn, C.G. Puntonet, J. Ortega, and J.M. Gόrriz (Spain)


Parallel matrix algorithms; Clusters of computers; Time series prediction; Artificial neural networks; Orthogonal transformations;


We analyze the parallel processing in clusters of comput ers of a prediction method based on the improvement of Radial Basis Function (RBF) neural networks using ma trix decomposition techniques such as the Singular Value Decomposition (SVD) and the QR-cp factorization. Par allel processing is required because of the extensive com putation found in those techniques, but the reward is ob tained in form of better prediction performance and less network complexity. This general prediction procedure (in the sequential version) was published in the technical liter ature previously, with a high degree of experimental suc cess. Parallelism is a convenient way to make this pre diction module available for inexpensive operation within decision-making contexts. We discuss two alternatives of concurrency: parallel implementation of the prediction procedure over the ScaLAPACK suite, and the formulation of another parallel routine customized to a higher degree for better performance.

Important Links:

Go Back