• Annual Conference of KIPS
• /
• 2004.11a
• /
• pp.493-496
• /
• 2004

본 연구에서는 신경망 학습의 중요한 평가 척도로써 고려될 수 있는 일반화 성능과 학습속도를 개선시키기 위한 방안으로써 초기 가중값과 학습률과 같은 주요 인자들을 이용한 신경망 학습 영향을 살펴본다. 특히 초기 가중값과 학습률을 고정시킨 후 새롭게 조정된 계수들을 점차적으로 변화시키는 새로운 인자 결합방법을 이용하여 신경망 학습량과 학습속도를 비교해 보고 계수조정을 통한 개선된 학습 영향을 살펴본다. 그리고 단순한 예제를 이용한 실증분석을 통하여 신경망 모형의 일반화 성능과 학습 속도 개선을 위한 각 인자들의 개별 효과와 결합 효과를 살펴보고 그 개선 방안을 제시한다.

• The KIPS Transactions:PartB
• /
• v.12B no.3 s.99
• /
• pp.343-348
• /
• 2005

The goal of this paper is to study the joint effect of factors of neural network teaming procedure. There are many factors, which may affect the generalization ability and teaming speed of neural networks, such as the initial values of weights, the learning rates, and the regularization coefficients. We will apply a constructive training algerian for neural network, then patterns are trained incrementally by considering them one by one. First, we will investigate the effect of these factors on generalization performance and learning speed. Based on these factors' effect, we will propose a joint method that simultaneously considers these three factors, and dynamically hue the learning rate and regularization coefficient. Then we will present the results of some experimental comparison among these kinds of methods in several simulated nonlinear data. Finally, we will draw conclusions and make plan for future work.

• The Korean Journal of Applied Statistics
• /
• v.29 no.3
• /
• pp.505-512
• /
• 2016

The paper is concerned with high frequency financial time series. A weighted hybrid volatility is suggested to compute daily volatilities based on high frequency data. Various realized volatility (RV) computations are reviewed and the weights are chosen by minimizing the differences between the hybrid volatility and the realized volatility. A high frequency time series of KOSPI200 index is illustrated via QLIKE and Theil-U statistics.

• The Korean Journal of Applied Statistics
• /
• v.26 no.5
• /
• pp.847-856
• /
• 2013

We propose a modified learning process for generalized neural network using a learning algorithm by Liu et al. (2001). We consider the effect of initial weights, training results and learning errors using a modified learning process. We employ an incremental training procedure where training patterns are learned systematically. Our algorithm starts with a single training pattern and a single hidden layer neuron. During the course of neural network training, we try to escape from the local minimum by using a weight scaling technique. We allow the network to grow by adding a hidden layer neuron only after several consecutive failed attempts to escape from a local minimum. Our optimization procedure tends to make the network reach the error tolerance with no or little training after the addition of a hidden layer neuron. Simulation results with suitable initial weights indicate that the present constructive algorithm can obtain neural networks very close to minimal structures and that convergence to a solution in neural network training can be guaranteed. We tested these algorithms extensively with small training sets.

• Survey Research
• /
• v.11 no.1
• /
• pp.123-144
• /
• 2010

In new sample design for Korea Housing Survey to research about housing policy, total strata are forty five because individual results of sixteen regions are estimated. The sample size is determined by sample errors of several variables which are the living area, family income, householder income, and living expenses. The sample size of each region is determined by relative standard error of existing result, and the strata sample size is to use the square root proportion allocation. Enumeration districts are sampled by the probability proportion to size systematic sampling in proportion to the enumeration district size, and the systemic sampling to use assortment characteristics. We considered a new apartment complex because of variation reflections which are rebuilder and redevelopment of houses. To get estimators of mean and variance, we used the design weighting, non-response adjusting, and post-stratification. In order to consider estimation efficiency, we calculate the design effect using estimators of variance.

• Journal of IKEEE
• /
• v.4 no.2 s.7
• /
• pp.249-256
• /
• 2000

In EIT(electrical impedance tomography), the internal resistivity(or conductivity) distribution of the unknown object is estimated using the boundary voltage data induced by different current patterns using various reconstruction algorithms. In this paper, we present a regularized modified Newton-Raphson(mNR) scheme which employs additional a priori information in the cost functional as soft constraint and the weighting matrices in the cost functional are selected based on the exponentially weighted least square criterion. The computer simulation for the 32 channels synthetic data shows that the reconstruction performance of the proposed scheme is improved compared to that of the conventional regularized mNR at the expense of slightly increased computational burden.

• The Korean Journal of Applied Statistics
• /
• v.26 no.3
• /
• pp.375-387
• /
• 2013

Unit non-response of surveys reduces the efficiency of the estimates and also causes non-response bias especially when there is large difference between respondents and non-respondents. Non-response weighting adjustments have usually been used to compensate for non-response. It is not easy to examine the non-response bias as well as to obtain information on the non-respondents in sample surveys. A household panel survey, called The Survey of Household Finances, was conducted in both 2010 and 2011. In this paper, we assume that non-response households in Wave 2 have strong non-response (non-cooperative) tendency. We classify those households into non-response households in Wave 1. Under this assumption, the characteristics of non-response households, the non-response bias and the effect of non-response adjustments are investigated.

• The Korean Journal of Applied Statistics
• /
• v.22 no.6
• /
• pp.1315-1329
• /
• 2009

The panel survey is need to have a more concern about a response due to a secession and non-response of a sample. And generally a population is not fixed and continuously changed. Thus, the rotation sample design can be used by the method replacing the panel research. This paper is the study of comparison to equal weight method, Duncan weight, Design weight method, weight share method in rotation sample design. More specifically, this paper compared variance estimators about the existing each method for the efficiency comparison, and to compare the precision using the relative efficiency gain by the Coefficient Variance(CV) after getting the design weight from the actual data.

• Annual Conference of KIPS
• /
• 2008.11a
• /
• pp.714-717
• /
• 2008

협업여과 추천기법에는 사용자 기반 협업여과와 아이템 기반 협업여과가 있으며, 절차는 유사도 측정, 이웃 선정, 예측값 생성 단계로 이루어진다. 유사도 측정 단계에는 유클리드 거리(Euclidean Distance), 코사인 유사도(Cosine Similarity), 피어슨 상관계수(Pearson Correlation Coefficient) 방법 등이 있고, 이웃 선정 단계에는 상관 한계치(Correlation-Threshold), 근접 N 이웃(Best-N-Neighbors) 방법 등이 있다. 마지막으로 예측값 생성 단계에는 단순평균(Simple Average), 가중합(Weighted Sum), 조정 가중합(Adjusted Weighted Sum) 등이 있다. 이처럼 협업여과 추천기법에는 다양한 기법들이 사용되고 있다. 따라서 본 논문에서는 사용자 기반 협업여과와 아이템 기반 협업여과 추천기법에 사용되는 유사도 측정 기법과 예측값 생성 기법의 최적화된 조합을 알아보기 위해 성능 실험 및 비교 분석을 하였다. 실험은 GroupLens의 MovieLens 데이터 셋을 활용하였고 MAE(Mean Absolute Error)값을 이용하여 추천기법을 비교 하였다. 실험을 통해 유사도 측정 기법과 예측값 생성 기법의 최적화된 조합을 찾을 수 있었고, 사용자 기반 협업여과와 아이템 기반 협업여과의 성능비교를 통해 아이템 기반 협업여과의 성능이 보다 우수했음을 확인 하였다.

• Spatial Information Research
• /
• v.19 no.2
• /
• pp.29-36
• /
• 2011

This study performed quantitative assessment of the performance of adjustment models by a-priori analysis of the statistics of the surface parameter estimates against modeling factors. Lidar, airborne imagery, and SAR imagery have been used to acquire the earth surface elevation, where the shape properties of the surface need to be determined through neighboring observations around target location. In this study, parameters which are selected to be estimated are elevation, slope, second order coefficient. In this study, several factors which are needed to be specified to compose adjustment models are classified into three types: mathematical functions, kernel sizes, and weighting types. Accordingly, a-priori standard deviations of the parameters are computed for varying adjustment models. Then their corresponding confidence regions for both the standard deviation of the estimate and the estimate itself are calculated in association with probability distributions. Thereafter, the resulting confidence regions are compared to each other against the factors constituting the adjustment models and the quantitative performance of adjustment models are ascertained.