• Journal of the Korean Institute of Telematics and Electronics B
• /
• v.31B no.12
• /
• pp.108-115
• /
• 1994

Most pattern classifiers have been designed based on the ML (Maximum Likelihood) training algorithm which is simple and relatively powerful. The ML training is an efficient algorithm to individually estimate the model parameters of each class under the assumption that all class models in a classifier are statistically independent. That assumption, however, is not valid in many real situations, which degrades the performance of the classifier. In this paper, we propose a minimum-error-rate training algorithm based on the MAP (Maximum a Posteriori) approach. The algorithm regards the normalized outputs of the classifier as estimates of the a posteriori probability, and tries to maximize those estimates. According to Bayes decision theory, the proposed algorithm satisfies the condition of minimum-error-rate classificatin. We apply this algorithm to NPM (Neural Prediction Model) for speech recognition, and derive new disrminative training algorithms. Experimental results on ten Korean digits recognition have shown the reduction of 37.5% of the number of recognition errors.

• Transactions of the Korean Society of Mechanical Engineers B
• /
• v.28 no.10
• /
• pp.1255-1263
• /
• 2004

An a priori test has been conducted for the dynamic mixed model which was generalized for the prediction of passive scalar field in a turbulent channel flow The results from a priori tests indicated that dynamic mixed model is capable of predicting both subgrid-scale heat flux and dissipation rather accurately. The success is attributed to the explicitly calculated resolved term incorporated into the model. The actual test of the model in a LES a posteriori showed that dynamic mixed model is superior to the widely used dynamic Smagorinsky model in the prediction of temperature statistics.

• The Journal of Korean Institute of Communications and Information Sciences
• /
• v.24 no.12B
• /
• pp.2447-2454
• /
• 1999

This paper proposes a N-APPEM(Nonlinear A Posteriori Probability Estimation Method) with a frame normalization method to conventional classification network to increase speech recognition ability. It also tests the recognition ability of the classification and prediction neural networks for the Korean isolated digits. From the experimental results, the prediction network with MLP(Multi-Layer Perceptron) achieves the highest recognition ability of 98.0%. The prediction requires very complicated networks increased linearly with the number of incoming speech categories. However, the classification network with the N-APPEM and the normalization improves the recognition ability up to 85.5% with a sin81e network, which is almost 12.0% improvement.

• Journal of The Korean Astronomical Society
• /
• v.41 no.4
• /
• pp.99-107
• /
• 2008

When a new counting experiment is proposed, it is crucial to predict whether the desired source signal will be detected, or how much observation time is required in order to detect the signal at a certain significance level. The concept of the a priori prediction of the detection limit in a newly proposed experiment should be distinguished from the a posteriori claim or decision whether a source signal was detected in an experiment already performed, and the calculation of statistical significance of a measured source signal. We formulate precise definitions of these concepts based on the statistical theory of hypothesis testing, and derive an approximate formula to estimate quickly the a priori detection limit of expected Poissonian source signals. A more accurate algorithm for calculating the detection limits in a counting experiment is also proposed. The formula and the proposed algorithm may be used for the estimation of required integration or observation time in proposals of new experiments. Applications include the calculation of integration time required for the detection of faint emission lines in a newly proposed spectroscopic observation, and the detection of faint sources in a new imaging observation. We apply the results to the calculation of observation time required to claim the detection of the surface thermal emission from neutron stars with two virtual instruments.

• Journal of Communications and Networks
• /
• v.10 no.1
• /
• pp.44-54
• /
• 2008

In this paper, an iterative multiple symbol differential detection for turbo coded differential unitary space-time modulation using a posteriori probability (APP) demodulator is investigated. Two approaches of different complexity based on linear prediction are presented to utilize the temporal correlation of fading for the APP demodulator. The first approach intends to take account of all possible previous symbols for linear prediction, thus requiring an increase of the number of trellis states of the APP demodulator. In contrast, the second approach applies Viterbi algorithm to assist the APP demodulator in estimating the previous symbols, hence allowing much reduced decoding complexity. These two approaches are found to provide a trade-off between performance and complexity. It is shown through simulation that both approaches can offer significant BER performance improvement over the conventional differential detection under both correlated slow and fast Rayleigh flat-fading channels. In addition, when comparing the first approach to a modified bit-interleaved turbo coded differential space-time modulation counterpart of comparable decoding complexity, the proposed decoding structure can offer performance gain over 3 dB at BER of $10^{-5}$.

• Journal of the Korean Data and Information Science Society
• /
• v.24 no.4
• /
• pp.857-865
• /
• 2013

Extreme rainfall causes heavy losses in human life and properties. Hence many works have been done to predict extreme rainfall by using extreme value distributions. In this study, we use a generalized extreme value distribution to derive the posterior predictive density with hierarchical Bayesian approach based on the data of Seoul area from 1973 to 2010. It becomes clear that the probability of the extreme rainfall is increasing for last 20 years in Seoul area and the model proposed works relatively well for both point prediction and predictive interval approach.

• Journal of Mechanical Science and Technology
• /
• v.19 no.3
• /
• pp.913-923
• /
• 2005

The dynamic mixed model (DMM) combined with a box filter of Zang et. al. (1993) has been generalized for passive scalar transport and applied to large eddy simulation of turbulent channel flows with Prandtl number up to 10. Results from a priori test showed that DMM is capable of predicting both subgrid-scale (SGS) scalar flux and dissipation rather accurately for the Prandtl numbers considered. This would suggest that the favorable feature of DMM, originally developed for the velocity field, works equally well for scalar transport problem. The validity of the DMM has also been tested a posteriori. The results of the large eddy simulation showed that DMM is superior to the dynamic Smagorinsky model in the prediction of scalar field and the model performance of DMM depends to a lesser degree on the ratio of test to grid filter widths, unlike in the a priori test.

• Journal of the Korean Society of Manufacturing Technology Engineers
• /
• v.9 no.5
• /
• pp.144-149
• /
• 2000

The demands for robotic and automatic system are continually increasing in manufacturing fields. There have been many studies to monitor and predict the system, but they have mainly focused upon measuring cutting force, and current of motor spindle, and upon using acoustic sensor, etc. In this study, time series sequence of cutting force was acquired by taking advantage of piezoelectric type tool dynamometer. Radial cutting force was obtained from it and was available for useful observation data. The parameter was estimated using PAA(parameter adaptation algorithm) from observation data. ARMA(auto regressive moving average) model was selected for system model and second order was decided according to parameter estimation. Uncorrelation test was also carried out to verify convergence of parameter.

• Proceedings of the Korean Society of Machine Tool Engineers Conference
• /
• 1999.10a
• /
• pp.40-45
• /
• 1999

The demands for robotic and automatic system are continually increasing in manufacturing fields. There have been many studies to monitor and predict the system, but they have mainly focused upon measuring cutting force, and current of motor spindle, and upon using acoustic sensor, etc.In this study, time series sequence of cutting force was acquired by taking advantage of piezoelectric type tool dynamometer. Radial cutting force was obtained from it and was available for useful observation data. The parameter was estimated using PAA (parameter adaptation algorithm) from observation data. ARMA(auto regressive moving average) model was selected for system model and second order was decided according to parameter estimation. Uncorrelation test was also carried out to verify convergence of parameter.

• Proceedings of the Acoustical Society of Korea Conference
• /
• spring
• /
• pp.75-78
• /
• 2000

본 논문은 자연스러운 합성음 생성을 위한 끊어 읽기 강도 예측에 관한 것으로, 문장에 대한 품사열이 주어졌을 때 Posteriori 확률을 최대화하는 끊어 읽기 강도를 비터비 디코딩으로 예측한다. 훈련용 데이터는 여성화자 1인이 발성한 2,100 문장이며, 음성 데이터로부터 휴지길이(pause)에 따라 끊어 읽기 강도를 2단계로 할당하고, 텍스트에서는 30개의 품사 태그 심볼을 이용하여 형태소분석 및 태깅을 수행하였다. 관측확률은 3개 연속하는 품사열이 발생할 확률로 하고 끊어 읽기 강도 천이확률은 bigram으로 했을 때, cross validation 방법으로 성능 평가를 수행하였다 평가결과, 훈련데이타에 대해서는 $89.7\%$, 테스트 데이터에 대해서는 $84.9\%$의 예측정확률을 보였다.