• Journal of KIISE:Software and Applications
• /
• v.30 no.3_4
• /
• pp.326-338
• /
• 2003

The objective of a neural network design and model selection is to construct an optimal network with a good generalization performance. However, training data include noises, and the number of training data is not sufficient, which results in the difference between the true probability distribution and the empirical one. The difference makes the teaming parameters to over-fit only to training data and to deviate from the true distribution of data, which is called the overfitting phenomenon. The overfilled neural network shows good approximations for the training data, but gives bad predictions to untrained new data. As the complexity of the neural network increases, this overfitting phenomenon also becomes more severe. In this paper, by taking statistical viewpoint, we proposed an integrative process for neural network design and model selection method in order to improve generalization performance. At first, by using the natural gradient learning with adaptive regularization, we try to obtain optimal parameters that are not overfilled to training data with fast convergence. By adopting the natural pruning to the obtained optimal parameters, we generate several candidates of network model with different sizes. Finally, we select an optimal model among candidate models based on the Bayesian Information Criteria. Through the computer simulation on benchmark problems, we confirm the generalization and structure optimization performance of the proposed integrative process of teaming and model selection.

• Journal of Digital Convergence
• /
• v.18 no.11
• /
• pp.259-266
• /
• 2020

Many developments have been steadily carried out by researchers with applying knowledge-based expert system or machine learning algorithms to the financial field. In particular, it is now common to perform knowledge based system trading in using stock prices. Recently, deep learning technologies have been applied to real fields of stock trading marketplace as GPU performance and large scaled data have been supported enough. Especially, LSTM has been tried to apply to stock price prediction because of its compatibility for time series data. In this paper, we implement stock price prediction using LSTM. In modeling of LSTM, we propose a fitness combination of model parameters and activation functions for best performance. Specifically, we propose suitable selection methods of initializers of weights and bias, regularizers to avoid over-fitting, activation functions and optimization methods. We also compare model performances according to the different selections of the above important modeling considering factors on the real-world stock price data of global major companies. Finally, our experimental work brings a fitness method of applying LSTM model to stock price prediction.

• Journal of Radiation Protection and Research
• /
• v.22 no.3
• /
• pp.153-160
• /
• 1997

The frequencies of ${\gamma}$-ray-induced micronuclei (MN) in cytokinesis-blocked (CB) lymphocytes at several doses were measured in three donors of human and C57BL/6 mice. Measurements performed after irradiation showed a dose-related increases in MN frequency in each of the donors studied. The relative sensitivity of mouse in spleen lymphocytes (SLs) compared with human peripheral blood lymphocytes (PBLs) was estimated by best fitting linear-quadratic model based on the radiation-induced MN data over the range from 0 cGy to 400 cGy. In the case of MN frequency with 0.2 per CB cell, the relative sensitivity of mouse SLs was 1.67. Compared with the radiation-induced MN formation in the PBLs of human, the SLs of mouse were more radiosensitive. Using this MN assay with human PBLs and mouse SLs, studies were performed to determine whether the water fraction of ginseng (Panax ginseng C.A.Meyer) against radiation-induced MN in human PBLs after in vitro irradiation (3Gy) and in SLs of C57BL/6 mice after in vivo irradiation (3Gy). The frequency of MN in human PBLs was reduced by water fraction of ginseng (0.5mg/ml of medium) both pre-and post treatment (p<0.0l) in vitro. In addition, the frequency of MN in mouse SLs was also reduced by pretreatment of ginseng (2mg/ml of drinking water for 7days) in vivo. The data suggested that the ginseng may reduce cell damage caused by ${\gamma}$-rays in vitro and in vivo. Further studies are needed to characterize better the protective nature of ginseng extract, its fractions and compounds.

• Journal of the Korean Institute of Landscape Architecture
• /
• v.39 no.6
• /
• pp.110-117
• /
• 2011

In order to study the growth models between the growth of Sedum takevimense and growth rate in soil with three types of mix ratios, this experiment was carried out on April 3rd, 2011. A nonlinearity regression analysis was performed using the Logistic and Gompertz models by SPSS. According to the study of growth models of Sedum takevimense, the process of growth and management methods after over-wintering were explicitly determined. According to the measured values, the growth in the soil of $P_1P_2V_1$ and $P_2P_1V_1$ was better than that of $P_1$. Particularly, the average length of Sedum takevimense in the soil of $P_1P_2V_1$ was about twice as great as that in the $P_1$. The fitness test of the two growth models was: The predicted value and measured value were separately compared and analysed, the average fitting precision $R^2$ of the Logistic models was 0.995, but the average $R^2$ of the Gompertz models was below 0.978, which showed that the Logistic models were better than the Gompertz models. The growth models also showed that the growth time of Sedum takevimense was divided into three: rapid, most rapid and slow. When managed in the rapid and the most rapid time, it will grow better.

• Journal of the Korean earth science society
• /
• v.22 no.6
• /
• pp.500-511
• /
• 2001

For the southeastern Korea aound the Yangsan fault we measured Q$_P^{-1}$ and Q$_S^{-1}$ simultaneously by using the extended coda-normalization method for seismograms registered at 9 stations deployed by KIGAM. We analyzed 707 seismograms of local earthquakes that occurred between December 1994 and February 2000. From seismograms, bandpass filtered traces were made by applying Butterworth filter with frequency-bands of 1${\sim}$2, 2${\sim}$4, 4${\sim}$8, 8${\sim}$16 and 16${\sim}$32 Hz. Estimated Q$_P^{-1}$ and Q$_S^{-1}$ values decrease from (7${\pm}$2)${\times}$10$^{-3}$ and (5${\pm}$4)${\times}$10$^{-4}$ at 1.5 Hz to (5${\pm}$4)${\times}$10$^{-3}$ and (5${\pm}$2)${\times}$10$^{-4}$ at 24 Hz, respectively. By fitting a power-law frequency dependent to estimated values over the whole stations, we obtained 0.009 (${\pm}$0.003)f$^{-1.05({\pm}0.14)$ for Q$_P^{-1}$ and 0.004 (${\pm}$0.001)f$^{-0.75({\pm}0.14)$) for Q$_S^{-1}$, where f is frequency in Hz.

• Journal of the Institute of Electronics Engineers of Korea SC
• /
• v.48 no.2
• /
• pp.47-55
• /
• 2011

Hemorrhagic shock is a cause of one third of death resulting from injury in the world. Early diagnosis of hemorrhagic shock makes it possible for physician to treat successfully. The objective of this paper was to select an optimal classifier model using physiological signals from rats measured during hemorrhagic experiment. This data set was used to train and predict survival rate using artificial neural network (ANN) and support vector machine (SVM). To avoid over-fitting, we chose the best classifier according to performance measured by a 10-fold cross validation method. As a result, we selected ANN having three hidden nodes with one hidden layer and SVM with Gaussian kernel function as trained prediction model, and the ANN showed 88.9 % of sensitivity, 96.7 % of specificity, 92.0 % of accuracy and the SVM provided 97.8 % of sensitivity, 95.0 % of specificity, 96.7 % of accuracy. Therefore, SVM was better than ANN for survival prediction.

• Journal of the Society of Naval Architects of Korea
• /
• v.54 no.5
• /
• pp.406-414
• /
• 2017

An experimental study in a recirculating water channel was carried out to investigate the effect of large coherent structures to the skin friction on a flat plate. Particle Image Velocimetry (PIV) technique was used to quantify characteristic features of coherent structures growing to the boundary layer. In the PIV measurement, it is difficult to calculate the friction velocity near the wall region due to laser deflection and uncertainty so that Clauser fitting method at the logarithmic region was adopted to compute the friction velocity and compared with the one directly measured by the dynamometer. With changing the free-stream velocity from 0.5 m/s to 1.0 m/s, the activity of coherent structures in the logarithmic region was increased over three times in terms of Reynolds stress. The flow field was separated by Variable Interval Time Averaging (VITA) technique into the weak and the strong structure case depending on the existence large coherent structures in order to validate its effectiveness. The stream-wise velocity fluctuation was scanned through at the boundary thickness whether it had a large deviation from background flow. With coherent structures connected from near-wall to the boundary layer, mean wall shear stress was higher than that of weak structure case. Proper Orthogonal Decomposition (POD) analysis was also applied to compare the energy budget between them at each free-stream velocity.

• Asian-Australasian Journal of Animal Sciences
• /
• v.30 no.3
• /
• pp.355-362
• /
• 2017

Objective: The use of locally produced forage (LPF) in cattle production has economic and environmental advantages over imported forage. The objective of this study was to characterize the nutritional value of LPF commonly used in Korea. Differences in ruminal fermentation characteristics were also examined for the LPF species commonly produced from two major production regions: Chungcheong and Jeolla. Methods: Ten LPF (five from each of the two regions) and six of the most widely used imported forages originating from North America were obtained at least three times throughout a year. Each forage species was pooled and analyzed for nutrient content using detailed chemical analysis. Ruminal fermentation characteristics were also determined by in vitro anaerobic incubations using strained rumen fluid for 0, 3, 6, 12, 24, and 48 h. At each incubation time, total gas, pH, ammonia, volatile fatty acid (VFA) concentrations, and neutral detergent fiber digestibility were measured. By fitting an exponential model, gas production kinetics were obtained. Results: Significant differences were found in the non-fiber carbohydrate (NFC) content among the forage species and the regions (p<0.01). No nutrient, other than NFC, showed significant differences among the regions. Crude protein, NFC, and acid detergent lignin significantly differed by forage species. The amount of acid detergent insoluble protein tended to differ among the forages. The forages produced in Chungcheong had a higher amount of NFC than that in Jeolla (p<0.05). There were differences in ruminal fermentation of LPF between the two regions and interactions between regions and forage species were also significant (p<0.05). The pH following a 48-h ruminal fermentation was lower in the forages from Chungcheong than from Jeolla (p<0.01), and total VFA concentration was higher in Chungcheong than in Jeolla (p = 0.05). This implies that fermentation was more active with the forages from Chungcheong than from Jeolla. Analysis of gas production profiles showed the rate of fermentation differed among forage species (p<0.05). Conclusion: The results of the present study showed that the nutritional values of some LPF (i.e., corn silage and Italian ryegrass) are comparable to those of imported forages widely used in Korea. This study also indicated that the nutritional value of LPF differs by origin, as well as by forage species. Detailed analyses of nutrient composition and digestion kinetics of LPF should be routinely employed to evaluate the correct nutritional value of LPF and to increase their use in the field.

• Journal of Korean Society for Geospatial Information Science
• /
• v.23 no.4
• /
• pp.43-48
• /
• 2015

We studied on pseudo-range correction(PRC) modeling in order to improve differential GNSS(DGNSS) accuracy. The PRC is the range correction information that provides improved location accuracy using DGNSS technique. The digital correction signal is typically broadcast over ground-based transmitters. Sometimes the degradation of the positioning accuracy caused by the loss of PRC signals, radio interference, etc. To prevent the degradation, in this paper, we have designed a PRC model through polynomial curve fitting and evaluated this model. We compared two quantities, estimations of PRC using model parameters and observations from the reference station. In the case of GPS, the average is 0.1m and RMSE is 1.3m. Most of GPS satellites have a bias error of less than ${\pm}1.0m$ and a RMSE within 3.0m. In the case of GLONASS, the average and the RMSE are 0.2m and 2.6m, respectively. Most of satellites have less than ${\pm}2.0m$ for a bias error and less than 3.0m for RMSE. These results show that the estimated value calculated by the model can be used effectively to maintain the accuracy of the user's location. However;it is needed for further work relating to the big difference between the two values at low elevation.

• Ecology and Resilient Infrastructure
• /
• v.3 no.4
• /
• pp.231-237
• /
• 2016

The purpose of this study is to analyze phosphate adsorption characteristic of the filter media applied in water purification technology. And it is also to observe the removal efficiency of the technology that was developed by combining the purification abilities of filter media and the vegetation mat. The filter media, Adphos, is the subject of this study. The result of adsorption test shows that $PO_4{^{3-}-P}$ adsorption characteristics of Adphos is determined by the Langmuir isotherm model fitting and the $q_m$ (maximum adsorption amount) value is calculated as 1.162 mg/g. To verify the removal efficiency of the purification technology, a lab-scale reactor, including a 400 mm depth of filter bed filled by Adphos, was manufactured. Yellow flag Iris (Iris pseudacours L.) was planted on the vegetation bed and the coir-roll was used to fix the root of vegetation. The reactor ran 8 hours per day over 71 days, and the results of filtration experiment show that average removal efficiency of SS, T-N and T-P were calculated as 94%, 41% and 64% respectively. With these results, it was proved that the purification technology is highly effective. And for the long-time use, a maintenance guide is also required.