An Application of Component-Wise Iterative Optimization to Feed-Forward Neural Networks
ICCS '07 Proceedings of the 7th international conference on Computational Science, Part II
A revisit to block and recursive least squares for parameter estimation
Computers and Electrical Engineering
| Hi-index | 35.68 |
Sliding window formulations of the fast QR and fast QR-lattice algorithms are presented. The derivations are based on the partial triangularization of raw data matrices. Three methods for window downdating are discussed: the method of plane hyperbolic rotations, the Chambers' method, and the LINPACK algorithm. A numerically ill-conditioned stationary signal and a speech signal are used in finite wordlength simulations of the full QR (nonfast), fast QR, and QR-lattice algorithms. All algorithms are observed to be numerically stable over billions of iterations for double-precision mantissas (53 bits), but as the number of bits is decreased in the mantissa, the algorithms exhibit divergent behavior. Hence, practically, the algorithms can de regarded as numerically stable for long wordlengths