Hiromu Gotanda

Dimensionally recursive order determination of linear discrete system

It is well known that the determinant ratio of the product moment matrix of observed input/output data gives a quick way of testing the order of a linear discrete system, and that the IPM (instrumental product moment) matrix is useful for eliminating the effect of the noise. This paper describes the dimensionally recursive algorithm for the determinant ratio associated with a new IPM matrix. This IPM matrix is constructed by using the lagged inputs themselves as the instrumental variables so that it would have a nesting structure. The algorithm is derived by taking advantage of its nesting structure. A particular quantity yielded naturally during the algorithm is shown to be a useful order test statistic, and a more robust order statistic is developed by using the parameter estimate given during the algorithm. By Monte Carlo simulations, the test procedures based on these statistics are confirmed to be valid in spite of the simple construction of the. IPM matrix even in the case where the noise is seriall...

Blind carrier frequency offset and channel estimation using ICA in QAM-OFDM systems

In this paper we propose a blind estimation of fading coefficients without using the pilot symbols in QAM-OFDM systems subjected to carrier frequency offset (CFO), where the phases of the fading coefficients are within ±7r/4. First, under a frequency selective fading channel, the inter-carrier interference (ICI) due to CFO is formulated in the framework of independent component analysis (ICA). Then the permutation and scale uncertainty inherent in ICA is resolved and CFO is estimated. Next, based on the CFO estimation results, the fading coefficients are estimated by use of the fact that symbol constellation at each sub-carrier is stretched/shrunk and rotated according to the value of its fading coefficient. Then, the unknown transmitted symbols are restored by use of the coefficient estimates. Finally, the validity of the proposed approach is confirmed from several simulations for 16QAM-OFDM.

Studies on Initialization for Multilayer Networks

Abstract This paper proposes an initialization of back propagation (BP) networks for pattern classification problems; the weights of hidden units are initialized so that hyperplanes should pass through the center of input pattern set, and those of the output layer are initialized to zero. Several simulation results confirm that the proposed initialization gives better convergence than the ordinary initialization that all the weights are initialized by uniform random values with zero mean.

Initialization of Back Propagation Algorithms for Multilayer Neural Networks

Abstract This paper examines the statistical properties of errors at the outset of the back propagation learning and derive a necessary condition for minimizing the errors that initial network outputs should take the same values to their target averages obtained a priori . It also proposes a new initialization incorporating a prioi information on the learning object into multilayer networks: of output layer units, the biases are initialized in terms of the target averages and the weights are initialized to zero. From simulation results, it has been confirmed that the initialization is useful to improve the learning convergence.

Noise Order Determination under Unknown Process Based on Covariance of Composite Noise and Output

Abstract In the present paper. we propose an efficient and robust method which determines not only the process order but also the noise order. The process order is evaluated by the statistics of instrumental residuals derived from the input instrumental product moment matrix which is so generated as to remove the noise effect. Although the noise covariance may be useful information to determine the noise order. the proper order can not be obtained under the situation required for the process identification from the estimate of the noise covariance because it is poor In accuracy. To determine the noise order. instead of the noise covariance estimate. We employ the cross-covariance estimate of the observed output and the composite noise which is a linear combination of the noise with the coefficient of the process parameter. taking notice of the fact that the cross-covariance is estimated accurately. The noise order is determined by the statistics yielded from the cross covariance. The proposed method is robust since it is not necessary to estimate the noise and exciter sequence as is the case for the determination method based on the extended least squares. Moreover. both statistics for the process and noise are efficiently computed by recursive algorithms.

Frequency domain blind channel estimation without phase ambiguity for QAM-OFDM systems

It is well known that, in the case of frequency domain blind channel estimation, the channel gains are correctly estimated but the channel phases are not necessarily estimated correctly: the channel estimates are often accompanied with phase ambiguity because the phase range for numerical calculation is restricted to (-π, π]. In this paper, a new frequency-domain blind channel estimation without phase ambiguity is proposed for QAM-OFDM systems of which transmitting and receiving stations are in LoS (Line of Sight) under the conditions: i) the absolute value of phase rotation is less than π/2 at all subcarriers, and ii) it is less than π/4 for at least one subcarrier among them. From several simulation results for 50 sub-carriers with 16 QAM using 2048 OFDM blocks, it is confirmed that, when the SNR (Signal-to-Noise Ratio) in baseband region is over 20[dB], the QAM-OFDM channel is estimated at high precision without phase ambiguity. The restored symbols resulting from the channel estimates give the bit-error-rate of about 10-3 or less at SNR≥20[dB] for the case that the ratio of the root-mean-square delay spread to the OFDM symbol length is 1/16.

Studies on Estimation of the Sources Number in Blind Source Separation Problems

ICA (Independent Component Analysis) can separate unknown source signals from their mixture signals without information on the transfer functions, provided that the sources are statistically independent. When the number of the source signals is equal to that of the observed signals, the original sources can be recovered except for indeterminacy of scale and permutation. However, the number of the sources is unknown in a real environment. In this paper, we propose an estimation method for the number of the sources based on the joint distribution of the observed signals under two-sensor configuration. From several simulation results, it is found that the number of the sources is coincident to that of peaks in the histogram of the distribution

Recognition of Similar Patterns by Multilayer Nets and Detection of Rotated Angle and Scale Ratio

For practical pattern recognition, it is required not only to recognize geometrically similar patterns but also to detect the difference of translation, rotation and size from their templates. This paper proposes a method to recognize the similar patterns by a multilayer net and then detect the difference on the common basis of well-known geometrical characteristics (center of gravity, angle of principal axis, and variance). It is found from experimental results that, with the proposed method, a small net can classify the similar patterns at a high recognition rate and detect their rotated angles and scale ratios with a high accuracy.

Necessary Conditions for Multilayer Nets having Solutions and Convergent Superiority of Bipolar Nets

Abstract This paper formulates a necessary condition for multilayer nets to have solutions by a set of normal vectors orthogonal to separation hyperplanes. Comparing the necessary condition to the distributions of normal vectors with the weights and biases initialized ordinarily by random numbers with zero mean, it is derived that bipolar nets are superior to unipolar nets in convergence of the back propagation learning initialized in such an ordinary manner.

Effects of the polarity of neural network units on back-propagation learning

This paper considers the neural network in which the initial values for the weights and the bias are given by random numbers as in usual cases. The results of BP learning in networks composed of unipolar units having an activity range from 0 to 1 and networks with bipolar units with a range from −0.5 to 0.5 are compared. When the input space is large, the separation hyperplane at the outset of learning passes near the center of the input space in the bipolar case, while that in the unipolar case passes near the vertex. Because of this property, the number of separation hyperplanes that effectively separate the input spaces of the layers during the updating or realization of the solution is larger in the bipolar case than in the unipolar case. The difference between the two becomes more remarkable with the increase of size. As a result of simulation, it is verified that the learning by the bipolar network gives better convergence for a wider range of initial values than the learning by the unipolar network when the network is large. It is shown also that the kinds of solution obtained by the unipolar network tend to be deviated.