A Simple Method of Calculating the Optimal Step Size in 4DVAR Technique

In four—dimensional variational data assimilation (4DVAR) technology, how to calculate the optimal step size is always a very important and indeed difficult task. It is directly related to the computational efficiency. In this research, a new method is proposed to calculate the optimal step size more effectively. Both nonlinear one—dimensional advection equation and two—dimensional inertial wave equation are used to test and compare the influence of different methods of the optimal step size calculations on the iteration steps, as well as the simulation results of 4DVAR processes. It is in evidence that the different methods have different influences. The calculating method is very important to determining whether the iteration is convergent or not and whether the convergence rate is large or small. If the calculating method of optimal step size is properly determined as demonstrated in this paper, then it can greatly enlarge the convergence rate and further greatly decrease the iteration steps. This research is meaningful since it not only makes an important improvement on 4DVAR theory, but also has useful practical application in improving the computational efficiency and saving the computational time. Key words 4DVAR - Optimal step size - Iterative convergence rate This work was supported by the National Natural Science Foundation under grants: 49735180 and 49675259, the “973 Project? CHERES(G 1998040907), the Project of Natural Science Foundation of Jiangsu Province(BK99020), and the Project Sponsored by the Scientific Research Foundation for the Returned Overseas Chinese Scholars.

An NLMS algorithm with optimized preparatory step-size parameters for SQUID-based MCG data processing

We present a new least-mean-square algorithm of adaptive filtering to improve the signal to noise ratio for magneto-cardiography data collected with high-temperature SQUID-based magnetometers. By frequently adjusting the adaptive parameter a go systematic optimum values in the course of the programmed procedure, the convergence is accelerated with a highest speed and the minimum steady-state error is obtained simultaneously. This algorithm may be applied to eliminate other non-steady relevant noises as well.

Speech Separation Based on Robust Independent Component Analysis

In this paper, we applied RobustICA to speech separation and made a comprehensive comparison to FastICA according to the separation results. Through a series of speech signal separation test, RobustICA reduced the separation time consumed by FastICA with higher stability, and speeches separated by RobustICA were proved to having lower separation errors. In the 14 groups of speech separation tests, separation time consumed by RobustICA was 3.185 s less than FastICA by nearly 68%. Separation errors of FastICA had a float between 0.004 and 0.02, while the errors of RobustlCA remained around 0.003. Furthermore, compared to FastICA, RobustlCA showed better separation robustness. Experimental results showed that RohustICA was successful to apply to the speech signal separation, and showed superiority to FastlCA in speech separation.

The Variational Assimilation Experiment of GPS Bending Angle

More and more new types of observational data provide many new opportunities for improving numerical weather forecasts. Among these, the GPS (Global Positioning System) bending angle is undoubtedly very important. There are many advantages of the GPS bending angle, such as high resolution, availability in all weather conditions, and global data coverage. Thus it is very valuable to assimilate GPS bending angle data into numerical weather models. This paper introduces how to obtain and assimilate the GPS bending angle. There are two methods of assimilation: the indirect method and direct method, and they are both introduced in this paper. During the minimizing process of variational assimilation, calculation efficiency is very important and the optimal step size greatly influences the algorithm efficiency. Based on the characteristics of the minimizing algorithm, we obtain an adaptive method for calculating the optimizing step suitable for all kinds of minimization algorithms through mathematical deduction. Finally, a numerical variational assimilation experiment is performed using the GPS bending angle data of 11 October 1995. The numerical results indicate the validity of the variational assimilation method and the adaptive method introduced here.

Dynamic parameterized learning for unsupervised domain adaptation

Unsupervised domain adaptation enables neural networks to transfer from a labeled source domain to an unlabeled target domain by learning domain-invariant representations.Recent approaches achieve this by directly matching the marginal distributions of these two domains.Most of them,however,ignore exploration of the dynamic trade-off between domain alignment and semantic discrimination learning,thus rendering them susceptible to the problems of negative transfer and outlier samples.To address these issues,we introduce the dynamic parameterized learning framework.First,by exploring domain-level semantic knowledge,the dynamic alignment parameter is proposed,to adaptively adjust the optimization steps of domain alignment and semantic discrimination learning.Besides,for obtaining semantic-discriminative and domain-invariant representations,we propose to align training trajectories on both source and target domains.Comprehensive experiments are conducted to validate the effectiveness of the proposed methods,and extensive comparisons are conducted on seven datasets of three visual tasks to demonstrate their practicability.