• 전기전자학회논문지
• /
• 제22권1호
• /
• pp.14-20
• /
• 2018

The purpose of training a convolutional neural network (CNN) is to obtain weight factors that give high classification accuracies. The initial values of hyper-parameters affect the training results, and it is important to train a CNN with a suitable hyper-parameter set of a learning rate, a batch size, the initialization of weight factors, and an optimizer. We investigate the effects of a single hyper-parameter while others are fixed in order to obtain a hyper-parameter set that gives higher classification accuracies and requires shorter training time using a proposed VGG-like CNN for training since the VGG is widely used. The CNN is trained for four datasets of CIFAR10, CIFAR100, GTSRB and DSDL-DB. The effects of the normalization and the data transformation for datasets are also investigated, and a training scheme using merged datasets is proposed.

• 스마트미디어저널
• /
• 제9권4호
• /
• pp.36-43
• /
• 2020

The Monte Carlo tree search (MCTS) is a popular method for implementing an intelligent game program. It has several hyper-parameters that require an optimization for showing the best performance. Due to the stochastic nature of the MCTS, the hyper-parameter optimization is difficult to solve. This paper uses the self-playing capability of the MCTS-based game program for optimizing the hyper-parameters. It seeks a winner path over the hyper-parameter space while performing the self-play. The top-q longest winners in the winner path compete for the final winner. The experiment using the 15-15-5 game (Omok in Korean name) showed a promising result.

• Journal of applied mathematics & informatics
• /
• 제34권3_4호
• /
• pp.319-330
• /
• 2016

The hyper-Zagreb index HM(G) of a simple graph G is defined as the sum of the terms (d_u+d_v)² over all edges uv of G, where d_u denotes the degree of the vertex u of G. In this paper, we present several upper and lower bounds on the hyper-Zagreb index in terms of some molecular structural parameters and relate this index to various well-known molecular descriptors.

• 전기전자학회논문지
• /
• 제24권4호
• /
• pp.969-974
• /
• 2020

Research into hyper parameter optimization (HPO) has recently revived with interest in models containing many hyper parameters, such as deep neural networks. In this paper, we introduce the most widely used HPO methods, such as grid search, random search, and Bayesian optimization, and investigate their characteristics through experiments. The MNIST data set is used to compare results in experiments to find the best method that can be used to achieve higher accuracy in a relatively short time simulation. The learning rate and weight decay have been chosen for this experiment because these are the commonly used parameters in this kind of experiment.

• International Journal of Internet, Broadcasting and Communication
• /
• 제15권2호
• /
• pp.240-247
• /
• 2023

In this paper, effective data management in industrial sites such as intelligent factories using time series data was studied. For effective management of time series data, variables considering the significance of the data were used, and hyper parameters calculated through LSTM AE were applied. We propose an optimized modeling considering the importance of each data section, and through this, outlier data of time series data can be efficiently processed. In the case of applying data significance and applying hyper parameters to which the research in this paper was applied, it was confirmed that the error rate was measured at 5.4%/4.8%/3.3%, and the significance of each data section and the significance of applying hyper parameters to optimize modeling were confirmed.

• 응용통계연구
• /
• 제24권1호
• /
• pp.177-183
• /
• 2011

Hidden Markov random eld(HMRF) is one of the most common model for image segmentation which is an important preprocessing in many imaging devices. The HMRF has unknown hyper-parameters on Markov random field to be estimated in segmenting testing images. However, in practice, due to computational complexity, it is often assumed to be a fixed constant. In this paper, we numerically show that the segmentation results very depending on the fixed hyper-parameter, and, if the parameter is misspecified, they further depend on the choice of the class-labelling algorithm. In contrast, the HMRF with estimated hyper-parameter provides consistent segmentation results regardless of the choice of class labelling and the estimation method. Thus, we recommend practitioners estimate the hyper-parameter even though it is computationally complex.

• Korean Journal of Radiology
• /
• 제23권4호
• /
• pp.446-454
• /
• 2022

Objective: To evaluate whether hyperoxia-induced ΔR₁ (hyperO₂ΔR₁) can accurately identify histological infarction in an acute cerebral stroke model. Materials and Methods: In 18 rats, MRI parameters, including hyperO₂ΔR₁, apparent diffusion coefficient (ADC), cerebral blood flow and volume, and ¹⁸F-fluorodeoxyglucose uptake on PET were measured 2.5, 4.5, and 6.5 hours after a 60-minutes occlusion of the right middle cerebral artery. Histological examination of the brain was performed immediately following the imaging studies. MRI and PET images were co-registered with digitized histological images. The ipsilateral hemisphere was divided into histological infarct (histological cell death), non-infarct ischemic (no cell death but ADC decrease), and nonischemic (no cell death or ADC decrease) areas for comparisons of imaging parameters. The levels of hyperO₂ΔR₁ and ADC were measured voxel-wise from the infarct core to the non-ischemic region. The correlation between areas of hyperO₂ΔR₁-derived infarction and histological cell death was evaluated. Results: HyperO₂ΔR₁ increased only in the infarct area (p ≤ 0.046) compared to the other areas. ADC decreased stepwise from non-ischemic to infarct areas (p = 0.002 at all time points). The other parameters did not show consistent differences among the three areas across the three time points. HyperO₂ΔR₁ sharply declined from the core to the border of the infarct areas, whereas there was no change within the non-infarct areas. A hyperO₂ΔR₁ value of 0.04 s^-1 was considered the criterion to identify histological infarction. ADC increased gradually from the infarct core to the periphery, without a pronounced difference at the border between the infarct and non-infarct areas. Areas of hyperO₂ΔR₁ higher than 0.04 s^-1 on MRI were strongly positively correlated with histological cell death (r = 0.862; p < 0.001). Conclusion: HyperO₂ΔR₁ may be used as an accurate and early (2.5 hours after onset) indicator of histological infarction in acute stroke.

• Journal of the Korean Data and Information Science Society
• /
• 제20권5호
• /
• pp.949-954
• /
• 2009

The autoregressive process is applied in this paper to kernel regression in order to infer nonlinear models for predicting responses. We propose a kernel method for the autoregressive data which estimates the mean function by kernel machines. We also present the model selection method which employs the cross validation techniques for choosing the hyper-parameters which affect the performance of kernel regression. Artificial and real examples are provided to indicate the usefulness of the proposed method for the estimation of mean function in the presence of autocorrelation between data.

• 한국컴퓨터정보학회논문지
• /
• 제29권4호
• /
• pp.31-38
• /
• 2024

본 연구는 온라인 원격교육에서 코딩 교육 활성화를 위해, 생성형 AI 기반의 학습 지원 도구개발에 필요한 하이퍼 파라미터 설정을 제안한다. 연구를 위해 세 가지 다른 학습 맥락에 따라 하이퍼 파라미터를 설정할 수 있는 실험 도구를 구현하고, 실험 도구를 통해 생성형 AI의 응답 품질을 평가하였다. 생성형 AI 자체의 기본 하이퍼 파라미터 설정을 유지한 실험은 대조군으로, 연구에서 설정한 하이퍼 파라미터를 사용한 실험은 실험군으로 하였다. 실험 결과, 첫 번째 학습맥락인 "학습 지원"에서는 실험군과 대조군 사이의 유의한 차이가 관찰되지 않았으나, 두 번째와 세 번째 학습 맥락인 "코드생성"과 "주석생성"에서는 실험군의 평가점수 평균이 대조군보다 각각 11.6% 포인트, 23% 포인트 높은 것으로 나타났다. 또한, system content에 응답이 학습 동기에 미칠 수 있는 영향을 제시하면 학습 정서를 고려한 응답이 생성되는 것이 관찰되었다.

• 제어로봇시스템학회:학술대회논문집
• /
• 제어로봇시스템학회 1997년도 한국자동제어학술회의논문집; 한국전력공사 서울연수원; 17-18 Oct. 1997
• /
• pp.442-445
• /
• 1997

We present a hybrid self-tuning method of fuzzy inference systems with hyper elliptic Gaussian membership functions using genetic algorithm(GA) and back-propagation algorithm. The proposed self-tuning method has two phases : one is the coarse tuning process based on GA and the other is the fine tuning process based on back-propagation. But the parameters which is obtained by a GA are near optimal solutions. In order to solve the problem in GA applications, it uses a back-propagation algorithm, which is one of learning algorithms in neural networks, to finely tune the parameters obtained by a GA. We provide Box-Jenkins time series to evaluate the advantage and effectiveness of the proposed approach and compare with the conventional method.