Mining Multivariate Time Series Models with Soft-Computing Techniques: A Coarse-Grained Parallel Computing Approach

This paper presents experimental results of a parallel implementation of a soft-computing algorithm for model discovery in multivariate time series, possibly with missing values. It uses a hybrid neural network with two different types of neurons trained with a non-traditional procedure. Models describing the multivariate time dependencies are encoded as binary strings representing neural networks, and evolved using genetic algorithms. The present paper studies its properties from an experimental point of view (using homogeneous and heterogeneous clusters) focussing on: i) the influence of missing values, ii) the factors controlling the parallel computation, and iii) the effectiveness of the time series prediction results. Results confirm that i) the algorithm possesses high tolerance to missing data, ii) Athon-based homogeneous clusters have higher throughput than Xeon-based homogeneous clusters, iii) an increase of the number of slaves reduces the processing time until communication overhead dominates (as expected), and iv) running the algorithm in parallel does not affect the RMS error (as expected). Even though much of this behavior could be qualitatively expected, appropriate tradeoffs between error and time were actually discovered, thereby enabling more effective, systematic, future uses of the system.