• Proceedings of the Korea Water Resources Association Conference
• /
• 2019.05a
• /
• pp.340-340
• /
• 2019

지구온난화로 유발되는 기후변동성이 증가함에 따라서 정확한 수문인자의 예측은 전 세계적으로 주요 관심사항이 되고 있다. 최근에는 고성능 컴퓨터 자원의 증가로 수문기상학 연구에서 동일한 학습량에 비하여 정확도의 향상이 뚜렷한 기계학습 구조를 활용하여 위성영상 기반의 대기예측, 태풍위치 추적 및 강수량 예측 등의 연구가 활발하게 진행되고 있다. 본 연구에는 기계학습 중 시계열 분석에 널리 활용되고 있는 순환신경망(Recurrent Neural Network, RNN) 기법의 대표적인 LSTM(Long Short-Term Memory) 네트워크를 이용하여 수문인자를 예측하였다. LSTM 네트워크는 가중치 및 메모리 요소에 대한 추가정보를 셀 상태에 저장하고 시계열의 길이 조정하여 모형의 탄력적 활용이 가능하다. LSTM 네트워크를 이용한 다양한 수문인자 예측결과 RMSE의 개선을 확인하였다. 따라서 본 연구를 통하여 개발된 기계학습을 통한 수문인자 예측기술은 권역별 수계별 홍수 및 가뭄대응 계획을 능동적으로 수립하는데 활용될 것으로 판단된다. 향후 연구에서는 LSTM의 입력영역을 Bayesian 추론기법을 활용하여 구성함으로 학습과정의 불확실성을 정량적으로 제어하고자 한다.

• The Journal of The Korea Institute of Intelligent Transport Systems
• /
• v.9 no.3
• /
• pp.73-84
• /
• 2010

This study presents a novel method to identify hazardous segments of freeway using global positioning system(GPS) based probe vehicle data. A variety of candidate contributing factors leading to higher potential of accident occurrence were extracted from the probe vehicle dataset. The research problem was defined as a classification problem, then a well-known classifier, bayesian neural network was adopted to solve the problem. A binary logistic regression technique was also used for selecting salient input variables. Test results showed that the proposed method is promising in extracting hazardous freeway sections. The outcome of this study will be effectively used for evaluating the safety of freeway sections and deriving countermeasures to prevent accidents.

• Journal of Korean Society of Transportation
• /
• v.24 no.2 s.88
• /
• pp.139-145
• /
• 2006

This paper develops pedestrian fatality models capable of producing the probability of pedestrian fatality in collision between vehicles and pedestrians. Probabilistic neural network (PNN) and binary logistic regression (BLR) ave employed in modeling pedestrian fatality pedestrian age, vehicle type, and collision speed obtained from reconstructing collected accidents are used as independent variables in fatality models. One of the nice features of this study is that an iterative sampling technique is used to construct various training and test datasets for the purpose of better performance comparison Statistical comparison considering the variation of model Performances is conducted. The results show that the PNN-based fatality model outperforms the BLR-based model. The models developed in this study that allow us to predict the pedestrian fatality would be useful tools for supporting the derivation of various safety Policies and technologies to enhance Pedestrian safety.

• Proceedings of the KOSOMBE Conference
• /
• v.1991 no.05
• /
• pp.65-70
• /
• 1991

A hybrid model which uses a probabilistic model and a MLP(multi layer perceptron) model for pattern recognition of EMG(electromyogram) signals is proposed in this paper. MLP model has problems which do not guarantee global minima of error due to learning method and have different approximation grade to bayesian probabilities due to different amounts and quality of training data, the number of hidden layers and hidden nodes, etc. Especially in the case of new test data which exclude design samples, the latter problem produces quite different results. The error probability of probabilistic model is closely related to the estimation error of the parameters used in the model and fidelity of assumtion. Generally, it is impossible to introduce the bayesian classifier to the probabilistic model of EMG signals because of unknown priori probabilities and is estimated by MLE(maximum likelihood estimate). In this paper we propose the method which get the MAP(maximum a posteriori probability) in the probabilistic model by estimating the priori probability distribution which minimize the error probability using the MLP. This method minimize the error probability of the probabilistic model as long as the realization of the MLP is optimal and approximate the minimum of error probability of each class of both models selectively. Alocating the reference coordinate of EMG signal to the outside of the body make it easy to suit to the applications which it is difficult to define and seperate using internal body coordinate. Simulation results show the benefit of the proposed model compared to use the MLP and the probabilistic model seperately.

• Journal of Korea Multimedia Society
• /
• v.6 no.1
• /
• pp.136-147
• /
• 2003

To achieve an optimal performance for a given problem, we need an integrative process of the parameter optimization via learning and the structure optimization via model selection. In this paper, we propose an efficient optimization method for improving generalization performance by considering the property of each sub-method and by combining them with common theoretical properties. First, weight parameters are optimized by natural gradient teaming with adaptive regularization, which uses a diverse error function. Second, the network structure is optimized by eliminating unnecessary parameters with natural pruning. Through iterating these processes, candidate models are constructed and evaluated based on the Bayesian Information Criterion so that an optimal one is finally selected. Through computational experiments on benchmark problems, we confirm the weight parameter and structure optimization performance of the proposed method.

• The KIPS Transactions:PartD
• /
• v.13D no.1 s.104
• /
• pp.133-138
• /
• 2006

Electric commerce of surpassing the traditional one appeared before the public and has currently led the change in the management of enterprises. To establish and maintain good relations with customers, electric commerce has various channels for customers that understand what they want to and suggest it to them. The bulletin board and e-mail among em are inbound information that enterprises can directly listen to customers' opinions and are different from other channels in characters. Enterprises can effectively manage the bulletin board and e-mail by understanding customers' ideas as many as possible and provide them with optimum answers. It is one of the important factors to improve the reliability of the notice board and e-mail as well as the whole electric commerce. Therefore this thesis researches into methods to classify various kinds of documents automatically in electric commerce; they are possible to solve existing problems of the bulletin board and e-mail, to operate effectively and to manage systematically. Moreover, it researches what the most suitable algorithm is in the automatic classification of Q&A documents by experiment the classifying performance of Naive Bayesian, TFIDF, Neural Network, k-NN

• The Journal of Information Systems
• /
• v.26 no.3
• /
• pp.47-61
• /
• 2017

Purpose Although automobile production has increased since the development of the Korean automobile industry, the number of customers who can purchase automobiles decreases relatively. Therefore, automobile companies need to develop strategies to attract customers and promote their repurchase behaviors. To this end, this paper analyzed customer data from a Korean automobile company using data mining techniques to derive repurchase strategies. Design/methodology/approach We conducted under-sampling to balance the collected data and generated 10 datasets. We then implemented prediction models by applying a decision tree, naive Bayesian, and artificial neural network algorithms to each of the datasets. As a result, we derived 10 patterns consisting of 11 variables affecting customers' decisions about repurchases from the decision tree algorithm, which yielded the best accuracy. Using the derived patterns, we proposed helpful strategies for improving repurchase rates. Findings From the top 10 repurchase patterns, we found that 1) repurchases in January are associated with a specific residential region, 2) repurchases in spring or autumn are associated with whether it is a weekend or not, 3) repurchases in summer are associated with whether the automobile is equipped with a sunroof or not, and 4) a customized promotion for a specific occupation increases the number of repurchases.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.9 no.8
• /
• pp.2948-2963
• /
• 2015

In the field of computer vision, visual surveillance systems have recently become an important research topic. Growth in this area is being driven by both the increase in the availability of inexpensive computing devices and image sensors as well as the general inefficiency of manual surveillance and monitoring. In particular, the ultimate goal for many visual surveillance systems is to provide automatic activity recognition for events at a given site. A higher level of understanding of these activities requires certain lower-level computer vision tasks to be performed. So in this paper, we propose an intelligent activity recognition model that uses a structure learning method and a classification method. The structure learning method is provided as a K2-learning algorithm that generates Bayesian networks of causal relationships between sensors for a given activity. The statistical characteristics of the sensor values and the topological characteristics of the generated graphs are learned for each activity, and then a neural network is designed to classify the current activity according to the features extracted from the multiple sensor values that have been collected. Finally, the proposed method is implemented and tested by using PETS2013 benchmark data.

• Journal of the Korean Institute of Telematics and Electronics B
• /
• v.28B no.10
• /
• pp.831-841
• /
• 1991

A combined model of probabilistic and MLP(multi layer perceptron) model is proposed for the pattern classification of EMG( electromyogram) signals. The MLP model has a problem of not guaranteeing the global minima of error and different quality of approximations to Bayesian probabilities. The probabilistic model is, however, closely related to the estimation error of model parameters and the fidelity of assumptions. A proper combination of these will reduce the effects of the problems and be robust to input variations. Proposed model is able to get the MAP(maximum a posteriori probability) in the probabilistic model by estimating a priori probability distribution using the MLP model adaptively. This method minimize the error probability of the probabilistic model as long as the realization of the MLP model is optimal, and this is a good combination of the probabilistic model and the MLP model for the usage of MLP model reliability. Simulation results show the benefit of the proposed model compared to use the Mlp and the probabilistic model seperately and the average calculation time fro classification is about 50ms in the case of combined motion using an IBM PC 25 MHz 386model.

• Journal of Internet Computing and Services
• /
• v.19 no.6
• /
• pp.101-112
• /
• 2018

A lot of manpower and budgets are being used to prevent fires, and only a small portion of the data generated during this process is used for disaster prevention activities. This study develops a prediction model of fire occurrence probability based on data mining in order to more actively use these data for disaster prevention activities. For this purpose, variables for predicting fire occurrence probability of various buildings were selected and data of construction administrative system, national fire information system, and Korea Fire Insurance Association were collected and integrated data set was constructed. After appropriate data cleansing and preprocessing, various data mining methodologies such as artificial neural network, decision trees, SVM, and Naive Bayesian were used to develop a prediction model of the fire occurrence probability of buildings. The most accurate model among the derived models is Linear SVM model which shows 68.42% as experimental data and 63.54% as verification data and it is the best model to predict fire occurrence probability of buildings. As this study develops the prediction model which uses only the set values of the specific ranges, future studies may explore more opportunites to use various setting values not shown in this study.