• 대기
• /
• 제32권4호
• /
• pp.367-379
• /
• 2022

In this paper, the evaluation of the performance of Korea Meteorological Administratio (KMA) Global Seasonal forecasting system version 6 (GloSea6) is presented by assessing the effects of larger ensemble size and carrying out the test using different initial conditions for hindcast in sub-seasonal to seasonal scales. The number of ensemble members increases from 3 to 7. The Ratio of Predictable Components (RPC) approaches the appropriate signal magnitude with increase of ensemble size. The improvement of annual variability is shown for all basic variables mainly in mid-high latitude. Over the East Asia region, there are enhancements especially in 500 hPa geopotential height and 850 hPa wind fields. It reveals possibility to improve the performance of East Asian monsoon. Also, the reliability tends to become better as the ensemble size increases in summer than winter. To assess the effects of using different initial conditions, the area-mean values of normalized bias and correlation coefficients are compared for each basic variable for hindcast according to the four initial dates. The results have better performance when the initial date closest to the forecasting time is used in summer. On the seasonal scale, it is better to use four initial dates, where the maximum size of the ensemble increases to 672, mainly in winter. As the use of larger ensemble size, therefore, it is most efficient to use two initial dates for 60-days prediction and four initial dates for 6-months prediction, similar to the current Time-Lagged ensemble method.

• 대기
• /
• 제31권5호
• /
• pp.511-523
• /
• 2021

This study explores the optimal ensemble size to improve the prediction performance of the Korea Meteorological Administration's operational climate prediction system, global seasonal forecast system version 5 (GloSea5). The GloSea5 produces an ensemble of hindcast data using the stochastic kinetic energy backscattering version2 (SKEB2) and timelagged ensemble. An experiment to increase the hindcast ensemble from 3 to 14 members for four initial dates was performed and the improvement and effect of the prediction performance considering Root Mean Square Error (RMSE), Anomaly Correlation Coefficient (ACC), ensemble spread, and Ratio of Predictable Components (RPC) were evaluated. As the ensemble size increased, the RMSE and ACC prediction performance improved and more significantly in the high variability area. In spread and RPC analysis, the prediction accuracy of the system improved as the ensemble size increased. The closer the initial date, the better the predictive performance. Results show that increasing the ensemble to an appropriate number considering the combination of initial times is efficient.

• 대기
• /
• 제20권3호
• /
• pp.287-306
• /
• 2010

An ensemble Kalman filter (EnKF) with Weather Research and Forecasting (WRF) Model is applied for Typhoon Wukong (200610) to investigate the performance of ensemble forecasts depending on experimental configurations of the EnKF. In addition, the ensemble sensitivity analysis is applied to the forecast and analysis ensembles generated in EnKF, to investigate the possibility of using the ensemble sensitivity analysis as the adaptive observation guidance. Various experimental configurations are tested by changing model error, ensemble size, assimilation time window, covariance relaxation, and covariance localization in EnKF. First of all, experiments using different physical parameterization scheme for each ensemble member show less root mean square error compared to those using single physics for all the forecast ensemble members, which implies that considering the model error is beneficial to get better forecasts. A larger number of ensembles are also beneficial than a smaller number of ensembles. For the assimilation time window, the experiment using less frequent window shows better results than that using more frequent window, which is associated with the availability of observational data in this study. Therefore, incorporating model error, larger ensemble size, and less frequent assimilation window into the EnKF is beneficial to get better prediction of Typhoon Wukong (200610). The covariance relaxation and localization are relatively less beneficial to the forecasts compared to those factors mentioned above. The ensemble sensitivity analysis shows that the sensitive regions for adaptive observations can be determined by the sensitivity of the forecast measure of interest to the initial ensembles. In addition, the sensitivities calculated by the ensemble sensitivity analysis can be explained by dynamical relationships established among wind, temperature, and pressure.

• Journal of the Korean Data and Information Science Society
• /
• 제27권1호
• /
• pp.255-264
• /
• 2016

Classification is a predictive modeling for a categorical target variable. Various classification ensemble methods, which predict with better accuracy by combining multiple classifiers, became a powerful machine learning and data mining paradigm. Well-known methodologies of classification ensemble are boosting, bagging and random forest. In this article, we assume that decision trees are used as classifiers in the ensemble. Further, we hypothesized that tree size affects classification accuracy. To study how the tree size in uences accuracy, we performed experiments using twenty-eight data sets. Then we compare the performances of ensemble algorithms; bagging, double-bagging, boosting and random forest, with different tree sizes in the experiment.

• 지능정보연구
• /
• 제16권1호
• /
• pp.93-116
• /
• 2010

불법현금융통 적발모형 개발에 앙상블 접근방법을 사용하였다. 불법현금융통은 국내 신용카드사의 손익에 영향을 미치며 최근 국제화되고 있음에도 불구하고 학문적인 접근이 이루어지지 않았다. 부정행위 적발모형(Fraud Detection Model, FDM)은 데이터 불균형 문제로 인하여 좋은 성능을 얻기 어려운데, 다수의 모형을 결합하는 앙상블이 대안으로 제시되어 왔다. 앙상블에 포함된 모형들의 다양성이 보장된다면 단일모형에 비해 더 좋은 성능을 보인다는 점은 이미 인정되고 있으며, 최근 연구 결과는 학습된 모든 기본모형들을 사용하는 것보다 적절한 기본모형들만 선택하여 앙상블에 포함시키는 것이 바람직하다는 것이다. 본 논문에서는 효과적인 불법현금융통 적발을 위하여 축소된 앙상블 기법을 사용하는데, 정확성과 다양성 척도를 사용하여 앙상블에 참여할 기본모형을 선택하는 것이다. 다양성은 앙상블을 구성하는 기본모형들 사이의 불일치 (Disagreement or Ambiguity)를 의미하는데, FDM에 내재된 데이터 불균형문제를 고려하여 두 가지 측면에 중점을 두었다. 첫째, 학습 자료의 추출 과정에서 다양성을 확보하기 위한 소수 범주의 과잉추출 방법과 적절한 훈련 방법에 대해 설명하였다. 둘째, 소수범주에 초점을 맞추어 기존의 다양성 척도를 효과적인 척도로 변형시키고, 전진추가법과 후진소거법의 동적 다양성 계산법을 도입하여 앙상블에 참여할 기본모형을 평가하였다. 실험에 사용된 학습 알고리즘은 신경망, 의사결정수와 로짓 회귀분석이었으며, 동질적 앙상블과 이질적 앙상블을 구성하여 성능평가를 하였다. 실험결과 불법현금융통 적발모형에 있어 축소된 앙상블은 모든 기본모형이 포함된 앙상블과 성능 차이가 없었다. 축소된 앙상블은 앙상블 구성의 복잡성을 감소시키고 구현을 용이하게 한다는 점에서 FDM에서도 유력한 모형 수립 접근방법이 될 수 있음을 보였다.

• 대기
• /
• 제28권4호
• /
• pp.509-520
• /
• 2018

The performance of the newly designed Pusan National University Coupled General Circulation Model (PNU CGCM) Ensemble Forecast System which produce 40 ensemble members for 12-month lead prediction is evaluated and analyzed in terms of boreal winter temperature over South Korea (S. Korea). The influence of ensemble size on prediction skill is examined with 40 ensemble members and the result shows that spreads of predictability are larger when the size of ensemble member is smaller. Moreover, it is suggested that more than 20 ensemble members are required for better prediction of statistically significant inter-annual variability of wintertime temperature over S. Korea. As for the ensemble average (ENS), it shows superior forecast skill compared to each ensemble member and has significant temporal correlation with Automated Surface Observing System (ASOS) temperature at 99% confidence level. In addition to forecast skill for inter-annual variability of wintertime temperature over S. Korea, winter climatology around East Asia and synoptic characteristics of warm (above normal) and cold (below normal) winters are reasonably captured by PNU CGCM. For the categorical forecast with $3{\times}3$ contingency table, the deterministic forecast generally shows better performance than probabilistic forecast except for warm winter (hit rate of probabilistic forecast: 71%). It is also found that, in case of concentrated distribution of 40 ensemble members to one category out of the three, the probabilistic forecast tends to have relatively high predictability. Meanwhile, in the case when the ensemble members distribute evenly throughout the categories, the predictability becomes lower in the probabilistic forecast.

• Communications for Statistical Applications and Methods
• /
• 제21권5호
• /
• pp.411-422
• /
• 2014

A classification ensemble method aggregates different classifiers obtained from training data to classify new data points. Voting algorithms are typical tools to summarize the outputs of each classifier in an ensemble. WAVE, proposed by Kim et al. (2011), is a new weight-adjusted voting algorithm for ensembles of classifiers with an optimal weight vector. In this study, when constructing an ensemble, we applied the WAVE algorithm on the double-bagging method (Hothorn and Lausen, 2003) to observe if any significant improvement can be achieved on performance. The results showed that double-bagging using WAVE algorithm performs better than other ensemble methods that employ plurality voting. In addition, double-bagging with WAVE algorithm is comparable with the random forest ensemble method when the ensemble size is large.

• 한국멀티미디어학회논문지
• /
• 제13권6호
• /
• pp.817-824
• /
• 2010

Linear Discriminant Analysis (LDA) has been successfully applied for dimension reduction in face recognition. However, LDA requires the transformation of a face image to a one-dimensional vector and this process can cause the correlation information among neighboring pixels to be disregarded. On the other hand, 2D-LDA uses 2D images directly without a transformation process and it has been shown to be superior to the traditional LDA. Nevertheless, there are some problems in 2D-LDA. First, it is difficult to determine the optimal number of feature vectors in a reduced dimensional space. Second, the size of rectangular windows used in 2D-LDA makes strong impacts on classification accuracies but there is no reliable way to determine an optimal window size. In this paper, we propose a new algorithm to overcome those problems in 2D-LDA. We adopt an ensemble approach which combines several classifiers obtained by utilizing various window sizes. And a practical method to determine the number of feature vectors is also presented. Experimental results demonstrate that the proposed method can overcome the difficulties with choosing an optimal window size and the number of feature vectors.

• 한국정보처리학회:학술대회논문집
• /
• 한국정보처리학회 2021년도 춘계학술발표대회
• /
• pp.373-376
• /
• 2021

Deep neural networks have achieved almost human-level results in various tasks and have become popular in the broad artificial intelligence domains. Uncertainty estimation is an on-demand task caused by the black-box point estimation behavior of deep learning. The deep ensemble provides increased accuracy and estimated uncertainty; however, linearly increasing the size makes the deep ensemble unfeasible for memory-intensive tasks. To address this problem, we used model pruning and quantization with a deep ensemble and analyzed the effect in the context of uncertainty metrics. We empirically showed that the ensemble members' disagreement increases with pruning, making models sparser by zeroing irrelevant parameters. Increased disagreement implies increased uncertainty, which helps in making more robust predictions. Accordingly, an energy-efficient compressed deep ensemble is appropriate for memory-intensive and uncertainty-aware tasks.

• 한국음향학회지
• /
• 제42권4호
• /
• pp.305-312
• /
• 2023

기포 크기 분포를 음파 감쇄 손실을 이용하여 역산하기 위해 Physics-Informed Neural Network(PINN)을 사용하였다. 역산에 사용되는 선형시스템을 풀기 위해 이미지 처리 분야에서 선형시스템 문제를 해결한 Adaptive Learned Iterative Shrinkage Thresholding Algorithm(Ada-LISTA)를 PINN의 신경망 구조로 이용하였다. 더 나아가, PINN의 손실함수에 선형시스템 기반의 정규항을 포함함으로써 PINN의 해가 기포 물리 법칙을 만족하여 더 높은 일반화 성능을 가지도록 하였다. 그리고 기포 추정값의 불확실성을 계산하기 위해 딥앙상블 기법을 이용하였다. 서로 다른 초기값을 갖는 20개의 Ada-LISTA는 같은 훈련데이터를 이용하여 학습되었다. 이 후 테스트시 훈련데이터와 다른 경향의 감쇄 손실을 입력으로 사용하여 기포 크기 분포를 추정하였고, 추정값과 이에 대한 불확실성을 20개 추정값의 평균과 분산으로 각각 구하였다. 그 결과 딥앙상블이 적용된 Ada-LISTA는 기존 볼록 최적화 기법인 CVX보다 기포 크기 분포를 역산하는데 더 우수한 성능을 보였다.