• The Journal of Bigdata
• /
• v.8 no.2
• /
• pp.15-25
• /
• 2023

Extreme learning machines (ELMs) are a major analytical method in various prediction fields. ELMs can accurately predict even if the data contains noise or is nonlinear by learning the complex patterns of time series data through optimal learning. This study presents the recent trends of machine learning models that are mainly studied as tools for analyzing online time series data, along with the application characteristics using existing algorithms. In order to efficiently learn large-scale online data that is continuously and explosively generated, it is necessary to have a learning technology that can perform well even in properties that can evolve in various ways. Therefore, this study examines a comprehensive overview of the latest machine learning models applied to big data in the field of time series prediction, discusses the general characteristics of the latest models that learn online data, which is one of the major challenges of machine learning for big data, and how efficiently they can learn and use online time series data for prediction, and proposes alternatives.

• Journal of the Korean Institute of Intelligent Systems
• /
• v.7 no.5
• /
• pp.1-7
• /
• 1997

In this paper, a neuro-contouring control scheme for the high precision machining of CNC machine tools is descrihed. The proposed control system consists of a conventional controller for each axis and an additional neuro-controller. For contouring control, the contour error must be computed during realtime motion, but generally the contour error for nonlinear contours is difficult to he directly computed. We, therefore, propose a new contour error model to approximate real error more exactly, and here we also introduce a cost function for better contouring performance and derive a learning law to adjust the weights of the neuro-controller. The derived learning law guarantees good contouring performance. Usefulness of the proposed control scheme is demonstrated hy computer simulations.

• Journal of the Korea Institute of Information and Communication Engineering
• /
• v.19 no.1
• /
• pp.106-112
• /
• 2015

In this paper, we propose an on-line signature recognition algorithm using fingertip point in the air from the depth image acquired by Kinect. We use ten statistical features for each X, Y, Z axis to react to changes in Shifting and Scaling of the signature trajectories in three-dimensional space. Artificial Neural Network is a machine learning algorithm used as a tool to solve the complex classification problem in pattern recognition. We implement the proposed algorithm to actual on-line signature recognition system. In experiment, we verify the proposed method is successful to classify 4 different on-line signatures.

• Proceedings of the Korean Society of Broadcast Engineers Conference
• /
• 2016.11a
• /
• pp.23-24
• /
• 2016

본 논문에서는 온라인 오디오 장르 분류의 성능을 비교 분석한다. 온라인 동작을 위해 1초 단위의 오디오 신호를 입력하여 music, speech, effect 중 하나의 장르로 판단한다. 학습 방법은 GMM과 심층 신경망을 사용하며, 특성은 MFCC와 스펙트로그램을 포함하는 네 가지 종류의 벡터를 사용한다. 각 성능을 비교 분석하여 장르 분류에 적합한 학습 방법과 특성 벡터를 확인한다.

• Journal of Broadcast Engineering
• /
• v.23 no.2
• /
• pp.186-196
• /
• 2018

In this paper, we propose an on-line tracking method using convolutional neural network (CNN) for tracking object. It is well known that a large number of training samples are needed to train the model offline. To solve this problem, we use an untrained model and update the model by collecting training samples online directly from the test sequences. While conventional methods have been used to learn models by training samples offline, we demonstrate that a small group of samples are sufficient for online object tracking. In addition, we define a loss function containing color information, and prevent the model from being trained by wrong training samples. Experiments validate that tracking performance is equivalent to four comparative methods or outperforms them.

• Journal of the Korean Institute of Intelligent Systems
• /
• v.14 no.4
• /
• pp.383-389
• /
• 2004

This paper proposes a method to provide intelligence for characters in fighting action games by using a neural network. Each action takes several time units in general fighting action games. Thus the results of a character's action are not exposed immediately but some time units later. To design a suitable neural network for such characters, it is very important to decide when the neural network is taught and which values are used to teach the neural network. The fitness of a character's action is determined according to the scores. For learning, the decision causing the score is identified, and then the neural network is taught by using the score change, the previous input and output values which were applied when the decision was fixed. To evaluate the performance of the proposed algorithm, many experiments are executed on a simple action game (but very similar to the actual fighting action games) environment. The results show that the intelligent character trained by the proposed algorithm outperforms random characters by 3.6 times at most. Thus we can conclude that the intelligent character properly reacts against the action of the opponent. The proposed method can be applied to various games in which characters confront each other, e.g. massively multiple online games.

• Proceedings of the Korean Information Science Society Conference
• /
• 2000.10b
• /
• pp.350-352
• /
• 2000

위성이 보내어오는 영상의 량은 인간이 일일이 실시간으로 검색할 수 없을 정도의 방대한 양이다. 그러므로 위성이 보내어오는 영상을 자동적으로 빠른 시간내에 분석하기 위하여 원패스로 성질이 유사한 영역을 묶어서 분류하는 알고리즘이 필요하다. 본 연구에서는 자기조직화 신경망(SOM)을 인공위성 영상을 원패스에 분할할 수 있도록 학습방법을 개선하였으며 개선된 SOM 알고리즘이 같은 원패스 알고리즘인 온라인 K-means과 비교하여 유효함을 알 수 있었다.

• Journal of KSNVE
• /
• v.8 no.6
• /
• pp.1181-1192
• /
• 1998

This paper describes the performance of a full-authority neural network-based fault tolerant system within a flight control system. This fault tolerant flight control system integrates sensor and actuator failure detection, identification, and accommodation (SFDIA and AFDIA), The first task is achieved by incorporating a main neural network (MNN) and a set of n decentralized neural networks (DNNs) to create a system for achieving fault tolerant capabilities for a system with n sensors assumed to be without physical redundancy The second scheme implements the same main neural network integrated with three neural network controllers (NNCs). The function of NNCs is to regain equilibrium and to compensate for the pitching, rolling. and yawing moments induced by the failure. Particular emphasis is placed in this study toward achieving an efficient integration between SFDIA and AFDIA without degradation of performance in terms of false alarm rates and incorrect failure identification. The results of the simulation with different actuator and sensor failures are presented and discussed.

• KIISE Transactions on Computing Practices
• /
• v.23 no.7
• /
• pp.440-445
• /
• 2017

Due to the advance of e-commerce systems, the number of people using online shopping and products has significantly increased. Therefore, the need for an accurate recommendation system is becoming increasingly more important. Recurrent neural network is a deep-learning algorithm that utilizes sequential information in training. In this paper, an evaluation is performed on the application of recurrent neural networks to recommendation systems. We evaluated three recurrent algorithms (RNN, LSTM and GRU) and three optimal algorithms(Adagrad, RMSProp and Adam) which are commonly used. In the experiments, we used the TensorFlow open source library produced by Google and e-commerce session data from RecSys Challenge 2015. The results using the optimal hyperparameters found in this study are compared with those of RecSys Challenge 2015 participants.

• Proceedings of the Korean Information Science Society Conference
• /
• 2011.06c
• /
• pp.292-294
• /
• 2011

전기자동차에 대한 연구가 진행됨에 따라 동력원으로 사용되는 대용량 리튬-폴리머 배터리의 운용과 관리에 대한 관심이 증가하고 있다. 다중 셀로 구성된 대용량 리튬-폴리머 배터리는 물리적 화학적 특성에 따라 충전시 셀간 전압 격차가 발생하게 된다. 셀간 전압차는 배터리 용량, 수명, 안정성에 부정적 영향을 주게 된다. 기존 연구들은 각 셀의 특성을 고려하지 않고 충전 결과를 바탕으로 동일한 밸런싱 방법을 적용시킴으로 효율성을 떨어트린다. 본 논문에서는 진화 신경망 기반의 지능형 셀 밸런싱 시스템을 제안한다. 배터리의 특성을 진화 신경망을 통해 학습시킴으로 각 셀 충전시 저항의 크기를 결정한다. 이를 통해 각 셀 특성을 고려한 사전 셀 밸런싱을 수행하였다. 제안하는 방법의 유용성을 입증하기 위해 카이스트 온라인 전기자동차에 장착 예정인 배터리 관리 시스템 기반 시뮬레이션을 수행하여 효과적인 셀 밸런싱이 가능함을 보였다.