• Journal of the Korea Computer Graphics Society
• /
• v.27 no.1
• /
• pp.9-17
• /
• 2021

In this paper, we present a simple and fast supervised learning framework based on model predictive control so as to learn motion controllers for a physic-based character to track given example motions. The proposed framework is composed of two components: training data generation and offline learning. Given an example motion, the former component stochastically controls the character motion with an optimal controller while repeatedly updating the controller for tracking the example motion through model predictive control over a time window from the current state of the character to a near future state. The repeated update of the optimal controller and the stochastic control make it possible to effectively explore various states that the character may have while mimicking the example motion and collect useful training data for supervised learning. Once all the training data is generated, the latter component normalizes the data to remove the disparity for magnitude and units inherent in the data and trains an artificial neural network with a simple architecture for a controller. The experimental results for walking and running motions demonstrate how effectively and fast the proposed framework produces physics-based motion controllers.

• Journal of Digital Convergence
• /
• v.20 no.4
• /
• pp.389-396
• /
• 2022

Deep Learning is based on a perceptron, and is currently being used in various fields such as image recognition, voice recognition, object detection, and drug development. Accordingly, a variety of learning algorithms have been proposed, and the number of neurons constituting a neural network varies greatly among researchers. This study analyzed the learning characteristics according to the number of neurons of the currently used SGD, momentum methods, AdaGrad, RMSProp, and Adam methods. To this end, a neural network was constructed with one input layer, three hidden layers, and one output layer. ReLU was applied to the activation function, cross entropy error (CEE) was applied to the loss function, and MNIST was used for the experimental dataset. As a result, it was concluded that the number of neurons 100-300, the algorithm Adam, and the number of learning (iteraction) 200 would be the most efficient in deep learning learning. This study will provide implications for the algorithm to be developed and the reference value of the number of neurons given new learning data in the future.

• Structural Monitoring and Maintenance
• /
• v.5 no.1
• /
• pp.51-65
• /
• 2018

A number of sensing techniques have been implemented for detecting defects in civil infrastructures instead of onsite human inspections in structural health monitoring. However, the issue of faults in sensors has not received much attention. This issue may lead to incorrect interpretation of data and false alarms. To overcome these challenges, this article presents a deep learning-based method with a new architecture of Stateful Long Short Term Memory Neural Networks (S-LSTM NN) for detecting sensor fault without going into details of the fault features. As LSTMs are capable of learning data features automatically, and the proposed method works without an accurate mathematical model. The detection of four types of sensor faults are studied in this paper. Non-stationary acceleration responses of a three-span continuous bridge when under operational conditions are studied. A deep network model is applied to the measured bridge data with estimation to detect the sensor fault. Another set of sensor output data is used to supervise the network parameters and backpropagation algorithm to fine tune the parameters to establish a deep self-coding network model. The response residuals between the true value and the predicted value of the deep S-LSTM network was statistically analyzed to determine the fault threshold of sensor. Experimental study with a cable-stayed bridge further indicated that the proposed method is robust in the detection of the sensor fault.

• Nuclear Engineering and Technology
• /
• v.56 no.4
• /
• pp.1472-1479
• /
• 2024

Cyber-Physical Energy Systems (CPESs) integrate cyber and hardware components to ensure a reliable and safe physical power production and supply. Renewable Energy Sources (RESs) add uncertainty to energy demand that can be dealt with flexible operation (e.g., load-following) of CPES; at the same time, scenarios that could result in severe consequences due to both component stochastic failures and aging of the cyber system of CPES (commonly overlooked) must be accounted for Operation & Maintenance (O&M) planning. In this paper, we make use of Deep Reinforcement Learning (DRL) to search for the optimal O&M strategy that, not only considers the actual system hardware components health conditions and their Remaining Useful Life (RUL), but also the possible accident scenarios caused by the failures and the aging of the hardware and the cyber components, respectively. The novelty of the work lies in embedding the cyber aging model into the CPES model of production planning and failure process; this model is used to help the RL agent, trained with Proximal Policy Optimization (PPO) and Imitation Learning (IL), finding the proper rejuvenation timing for the cyber system accounting for the uncertainty of the cyber system aging process. An application is provided, with regards to the Advanced Lead-cooled Fast Reactor European Demonstrator (ALFRED).

• Korean Journal of Remote Sensing
• /
• v.38 no.6_1
• /
• pp.1329-1341
• /
• 2022

Destroying the marine environment and marine ecosystem and causing marine accidents, marine debris is generated every year, and among them, submerged marine debris is difficult to identify and collect because it is on the seabed. Therefore, deep-learning-based semantic segmentation was experimented on waste fish nets and waste ropes using underwater images to identify efficient collection and distribution. For segmentation, a high-resolution network (HRNet), a state-of-the-art deep learning technique, was used, and the performance of each optimizer was compared. In the segmentation result fish net, F1 score=(86.46%, 86.20%, 85.29%), IoU=(76.15%, 75.74%, 74.36%), For the rope F1 score=(80.49%, 80.48%, 77.86%), IoU=(67.35%, 67.33%, 63.75%) in the order of adaptive moment estimation (Adam), Momentum, and stochastic gradient descent (SGD). Adam's results were the highest in both fish net and rope. Through the research results, the evaluation of segmentation performance for each optimizer and the possibility of segmentation of marine debris in the latest deep learning technique were confirmed. Accordingly, it is judged that by applying the latest deep learning technique to the identification of submerged marine debris through underwater images, it will be helpful in estimating the distribution of marine sedimentation debris through more accurate and efficient identification than identification through the naked eye.

• Management & Information Systems Review
• /
• v.2
• /
• pp.75-93
• /
• 1998

In this paper, we propose a new algorithm(N_BP) to be capable of overcoming limitations of the traditional backpropagation(O_BP). The N_BP is based on the method of conjugate gradients and calculates learning parameters through the line search which may be characterized by order statistics and golden section. Experimental results showed that the N_BP was definitely superior to the O_BP with and without a stochastic term in terms of accuracy and rate of convergence and might surmount the problem of local minima. Furthermore, they confirmed us that the stagnant phenomenon of learning in the O_BP resulted from the limitations of its algorithm in itself and that unessential approaches would never cured it of this phenomenon.

• Proceedings of the Korean Institute of Intelligent Systems Conference
• /
• 2007.04a
• /
• pp.33-36
• /
• 2007

최근에 인공지능 분야에서는, 국내외적으로 강화학습(reinforcement learning)에 관한 관심이 크게 증폭되고 있다. 강화학습의 최근 경향을 살펴보면, 크게 가치함수를 직접 활용하는 방법(value function-based methods), 제어 전략에 대한 탐색을 활용하는 방법(policy search methods), 그리고 액터-크리틱 방법(actor-critic methods)의 세가지 방향으로 발전하고 있음을 알 수 있다. 본 논문에서는 이중 세 번째 부류인 액터-크리틱 방법 중 NAC(natural actor-critic) 기법의 한 종류인 RLS-NAC(recursive least-squares based natural actor-critic) 알고리즘을 다양한 트레이스 감쇠계수를 사용하여 연속제어입력(real-valued control inputs)으로 제어되는 Kimura의 기는 로봇에 대해 적용해보고, 그 성능을 기존의 SGA(stochastic gradient ascent) 알고리즘을 이용하여 학습한 경우와 비교해보도록 한다.

• Korean Journal of Cognitive Science
• /
• v.10 no.1
• /
• pp.13-24
• /
• 1999

In this paper, we propose a classifier based on modular networks using an unsupervised learning method. The structure of each module is designed through stochastic analysis of input data and each module classifier data independently. The result of independent classification of each module and a measure of the nearest distance are integrated during the final data classification phase to allow more precise c classification. Computation time is decreased by deleting modules that have been classified to be incorrect during the final classification phase. Using this method. a neural network sharing the best performance was implemented without considering. lots of of variables which can affect the performance of the neural network.

• Communications for Statistical Applications and Methods
• /
• v.8 no.3
• /
• pp.625-632
• /
• 2001

Principal component analysis(PCA) is a multivariate technique falling under the general title of factor analysis. The purpose of PCA is to Identify the dependence structure behind a multivariate stochastic observation In order to obtain a compact description of it. In engineering field PCA is utilized mainly (or data compression and restoration. In this paper we propose a new robust Hebbian algorithm for robust PCA. This algorithm is based on a hyperbolic tangent function due to Hampel ef al.(1989) which is known to be robust in Statistics. We do two experiments to investigate the performance of the new robust Hebbian learning algorithm for robust PCA.

• Journal of Institute of Control, Robotics and Systems
• /
• v.13 no.4
• /
• pp.286-295
• /
• 2007

This paper presents an intelligent PID control for stochastic systems with nonstationary nature. We optimally determine parameters of a PID controller through learning algorithm and propose an online PID control to compensate system errors possibly occurred in realtime implementations. A dynamic Bayesian network (DBN) model for system errors is additionally explored for making decision about whether an online control is carried out or not in practice. We apply our control approach to traffic control of Transmission Control Protocol (TCP) networks and demonstrate its superior performance comparing to a fixed PID from computer simulations.