• ETRI Journal
• /
• v.15 no.2
• /
• pp.35-51
• /
• 1993

This paper presents function approximation based on nonparametric estimation. As an estimation model of function approximation, a three layered network composed of input, hidden and output layers is considered. The input and output layers have linear activation units while the hidden layer has nonlinear activation units or kernel functions which have the characteristics of bounds and locality. Using this type of network, a many-to-one function is synthesized over the domain of the input space by a number of kernel functions. In this network, we have to estimate the necessary number of kernel functions as well as the parameters associated with kernel functions. For this purpose, a new method of parameter estimation in which linear learning rule is applied between hidden and output layers while nonlinear (piecewise-linear) learning rule is applied between input and hidden layers, is considered. The linear learning rule updates the output weights between hidden and output layers based on the Linear Minimization of Mean Square Error (LMMSE) sense in the space of kernel functions while the nonlinear learning rule updates the parameters of kernel functions based on the gradient of the actual output of network with respect to the parameters (especially, the shape) of kernel functions. This approach of parameter adaptation provides near optimal values of the parameters associated with kernel functions in the sense of minimizing mean square error. As a result, the suggested nonparametric estimation provides an efficient way of function approximation from the view point of the number of kernel functions as well as learning speed.

• Journal of the Korean Institute of Telematics and Electronics S
• /
• v.34S no.1
• /
• pp.72-79
• /
• 1997

Most open magnetic resonance imaging systems have used the planar gradient coils whose inductances were minimized through the magnetic energy minimization procedure in the spatial frequency domain. Though the planar gradient coils have smaller inductance than conventional gradient coils, the planar gradient coils often suffer from their poor magnetic field linearity. Scaling the spatial frequencies of the current density function designed by the magnetic energy minimization, magnetic field linearity of the planar gradient coils can be greatly improved with small sacrifice of gradient coil inductance. We have found that the figure of merit of the planar gradient coils, defined by the gradient strength divided by the linearity error and the inductance, can be improved by proposed technique.

• Journal of the Korea Institute of Information and Communication Engineering
• /
• v.9 no.7
• /
• pp.1569-1574
• /
• 2005

In this paper, I propose an intelligent credit rating system using a bankruptcy prediction model based on support vector machines (SVMs). SVMs are promising methods because they use a risk function consisting of the empirical error and a regularized term which is derived from the structural risk minimization principle. This study examines the feasibility of applying SVM in Predicting corporate bankruptcies by comparing it with other data mining techniques. In addition. this study presents architecture and prototype of intelligeht credit rating systems based on SVM models.

• The Journal of Korean Institute of Communications and Information Sciences
• /
• v.35 no.8A
• /
• pp.760-767
• /
• 2010

For complex channel blind equalization, this study presents the performance and characteristics of two complex blind information theoretic learning algorithms (ITL) which are based on minimization of Euclidian distance (ED) between probability density functions compared to constant modulus algorithm which is based on mean squared error (MSE) criterion. The complex-valued ED algorithm employing constant modulus error and the complex-valued ED algorithm using a self-generated symbol set are analyzed to have the fact that the cost function of the latter forces the output signal to have correct symbol values and compensate amplitude and phase distortion simultaneously without any phase compensation process. Simulation results through MSE convergence and constellation comparison for severely distorted complex channels show significantly enhanced performance of symbol-point concentration with no phase rotation.

• International Journal of Advanced Culture Technology
• /
• v.5 no.2
• /
• pp.74-81
• /
• 2017

Recently, there are many companies that use systems based on artificial intelligence. The accuracy of artificial intelligence depends on the amount of learning data and the appropriate algorithm. However, it is not easy to obtain learning data with a large number of entity. Less data set have large generalization errors due to overfitting. In order to minimize this generalization error, this study proposed DGA(Dropout Genetic Algorithm) which can expect relatively high accuracy even though data with a less data set is applied to machine learning based genetic algorithm to deep learning based dropout. The idea of this paper is to determine the active state of the nodes. Using Gradient about loss function, A new fitness function is defined. Proposed Algorithm DGA is supplementing stochastic inconsistency about Dropout. Also DGA solved problem by the complexity of the fitness function and expression range of the model about Genetic Algorithm As a result of experiments using MNIST data proposed algorithm accuracy is 75.3%. Using only Dropout algorithm accuracy is 41.4%. It is shown that DGA is better than using only dropout.

• Journal of the Korean Society for Precision Engineering
• /
• v.23 no.9 s.186
• /
• pp.76-83
• /
• 2006

Machining accuracy is closely related with tool deflection induced by cutting forces. In this research, cutting forces and tool deflection in end milling are expressed as a closed form of tool rotational angle and cutting conditions. The discrete cutting fores caused by periodic tool entry and exit are represented as a continuous function using the Fourier series expansion. Tool deflection is predicted by direct integration of the distributed loads on cutting edges. Cutting conditions, tool geometry, run-outs and the stiffness of tool clamping part are considered together far cutting forces and tool deflection estimation. Compared with numerical methods, the presented method has advantages in prediction time reduction and the effects of feeding and run-outs on cutting forces and tool deflection can be analyzed quantitatively. This research can be effectively used in real time machining error estimation and cutting condition selection for error minimization since the form accuracy is easily predicted from tool deflection curve.

• ETRI Journal
• /
• v.35 no.2
• /
• pp.200-206
• /
• 2013

Most of the literature on compressed sensing has not paid enough attention to scenarios in which the number of acquired measurements is insufficient to satisfy minimal exact reconstruction requirements. In practice, encountering such scenarios is highly likely, either intentionally or unintentionally, that is, due to high sensing cost or to the lack of knowledge of signal properties. We analyze signal reconstruction performance in this setting. The main result is an expression of the reconstruction error as a function of the number of acquired measurements.

• Proceedings of the KIEE Conference
• /
• 1995.07b
• /
• pp.758-760
• /
• 1995

The purpose of noise cancellation is to estimating signals corrupted by additive noise or interference. In this paper, an adaptive noise canceller is built from a Walsh filter with a new adaptive algorithm. The Walsh filter consists of a Walsh function. Since the Walsh functions are either even or odd functions, the covariance matrix in the tap gain adjustment algorithm can be reduced to a simple form. In this paper, minimization of the mean squre error is accomplished by a proposed adaptive algorithm. The conventional adaptation techniques use a fixed time constant convergence factor by trial and error methods. In this paper, a convergence factor is obtained that is tailored for each adaptive filter coefficient and is updated at each block iteration.

• Journal of Internet Computing and Services
• /
• v.19 no.3
• /
• pp.49-56
• /
• 2018

Euclidean distance (ED) between error distribution and Dirac delta function has been used as an efficient performance criterion in impulsive noise environmentsdue to the outlier-cutting effect of Gaussian kernel for error signal. The gradient of ED for its minimization has two components; $A_k$ for kernel function of error pairs and the other $B_k$ for kernel function of errors. In this paper, it is analyzed that the first component is to govern gathering close together error samples, and the other one $B_k$ is to conduct error-sample concentration on zero. Based upon this analysis, it is proposed to normalize $A_k$ and $B_k$ with power of inputs which are modified by kernelled error pairs or errors for the purpose of reinforcing their roles of narrowing error-gap and drawing error samples to zero. Through comparison of fluctuation of steady state MSE and value of minimum MSE in the results of simulation of multipath equalization under impulsive noise, their roles and efficiency of the proposed normalization method are verified.

• Journal of Power Electronics
• /
• v.18 no.5
• /
• pp.1398-1408
• /
• 2018

In order to further improve the steady-state control performance of model predictive torque control (MPTC), a double-vector-based model predictive torque control without a weighting factor is proposed in this paper. The extended voltage vectors synthesized by two basic voltage vectors are used to increase the number of feasible voltage vectors. Therefore, the control precision of the torque and the stator flux along with the steady-state performance can be improved. To avoid testing all of the feasible voltage vectors, the solution of deadbeat torque control is calculated to predict the reference voltage vector. Thus, the candidate voltage vectors, which need to be evaluated by a cost function, can be reduced based on the sector position of the predicted reference voltage vector. Furthermore, a cost function, which only includes a reference voltage tracking error, is designed to eliminate the weighting factor. Moreover, two voltage vectors are applied during one control period, and their durations are calculated based on the principle of reference voltage tracking error minimization. Finally, the proposed method is tested by simulations and experiments.