The threat of North Korea's long-range firepower is recognized as a typical asymmetric threat, and South Korea is prioritizing the development of a Korean-style missile defense system to defend against it. To address this, previous research modeled North Korean long-range artillery attacks as a Markov Decision Process (MDP) and used Approximate Dynamic Programming as an algorithm for missile defense, but due to its limitations, there is an intention to apply deep reinforcement learning techniques that incorporate deep learning. In this paper, we aim to develop a missile defense system algorithm by applying a modified DQN with multi-agent-based deep reinforcement learning techniques. Through this, we have researched to ensure an efficient missile defense system can be implemented considering the style of attacks in recent wars, such as how effectively it can respond to enemy missile attacks, and have proven that the results learned through deep reinforcement learning show superior outcomes.

증산은 적정 관수 관리에 중요한 역할을 하므로 수분 스트레스에 취약한 토마토와 같은 작물의 관개 수요에 대한 지식이 필요하다. 관수량을 결정하는 한 가지 방법은 증산량을 측정하는 것인데, 이는 환경이나 생육 수준의 영향을 받는다. 본 연구는 분단위 데이터를 통해 수학적 모델과 딥러닝 모델을 활용하여 토마토의 증발량을 추정하 고 적합한 모델을 찾는 것을 목표로 한다. 라이시미터 데이터는 1분 간격으로 배지무게 변화를 측정함으로써 증산 량을 직접 측정했다. 피어슨 상관관계는 관찰된 환경 변수가 작물 증산과 유의미한 상관관계가 있음을 보여주었다. 온실온도와 태양복사는 증산량과 양의 상관관계를 보인 반면, 상대습도는 음의 상관관계를 보였다. 다중 선형 회귀 (MLR), 다항 회귀 모델, 인공 신경망(ANN), Long short-term memory(LSTM), Gated Recurrent Unit(GRU) 모델을 구 축하고 정확도를 비교했다. 모든 모델은 테스트 데이터 세트에서 0.770－0.948 범위의 R2 값과 0.495mm/min－ 1.038mm/min의 RMSE로 증산을 잠재적으로 추정하였다. 딥러닝 모델은 수학적 모델보다 성능이 뛰어났다. GRU 는 0.948의 R2 및 0.495mm/min의 RMSE로 테스트 데이터에서 최고의 성능을 보여주었다. LSTM과 ANN은 R2 값이 각각 0.946과 0.944, RMSE가 각각 0.504m/min과 0.511로 그 뒤를 이었다. GRU 모델은 단기 예측에서 우수한 성능 을 보였고 LSTM은 장기 예측에서 우수한 성능을 보였지만 대규모 데이터 셋을 사용한 추가 검증이 필요하다. FAO56 Penman-Monteith(PM) 방정식과 비교하여 PM은 MLR 및 다항식 모델 2차 및 3차보다 RMSE가 0.598mm/min으로 낮지만 분단위 증산의 변동성을 포착하는 데 있어 모든 모델 중에서 가장 성능이 낮다. 따라서 본 연구 결과는 온실 내 토마토 증산을 단기적으로 추정하기 위해 GRU 및 LSTM 모델을 권장한다.

전산유체역학을 사용하는 일반적인 선박의 저항성능 평가는 많은 시간과 비용이 필요하며, 이를 줄이기 위한 다양한 방법이 연구되고 있다. 선박의 주요 치수나 단면을 이용하는 기존의 방법들은 선형에 크게 좌우되는 저항성능을 추정하는데 한계가 있다. 본 논 문에서는 선형 격자의 기하학적 정보를 입력으로 선체 표면의 저항성능을 빠르게 추정할 수 있는 심층신경망 모델을 제안한다. Perceiver IO 기반의 제안하는 심층신경망 모델은 시간 단계별로 계산이 필요한 전산유체역학 기법과 달리 바로 저항성능 추정이 가능하며, 저속비 대선의 일종인 50K 탱커 선박을 대상으로 한 데이터집합에서 평균 1% 미만의 오차로 저항성능을 추정하는 결과를 보인다.

Deep convolutional network is a deep learning approach to optimize image recognition. This study aimed to apply DCNN to the reading of mandibular cortical thinning in digital panoramic radiographs. Digital panoramic radiographs of 1,268 female dental patients (age 45.2 ± 21.1yrs) were used in the reading of the mandibular cortical bone by two maxillofacial radiologists. Among the subjects, 535 normal subject’s panoramic radiographs (age 28.6 ±7.4 yrs) and 533 those of osteoporosis pationts (age 72.1 ± 8.7 yrs) with mandibular cortical thinning were used for training DCNN. In the testing of mandibular cortical thinning, 100 panoramic radiographs of normal subjects (age 26.6 ± 4.5 yrs) and 100 mandibular cortical thinning (age 72.5 ± 7.2 yrs) were used. The sensitive area of DCNN to mandibular cortical thinning was investigated by occluding analysis. The readings of DCNN were compared by two maxillofacial radiologists. DCNN showed 97.5% accuracy, 96% sensitivity, and 99% specificity in reading mandibular cortical thinning. DCNN was sensitively responded on the cancellous and cortical bone of the mandibular inferior area. DCNN was effective in diagnosing mandibular cortical thinning.

Control performance of a smart tuned mass damper (TMD) mainly depends on control algorithms. A lot of control strategies have been proposed for semi-active control devices. Recently, machine learning begins to be applied to development of vibration control algorithm. In this study, a reinforcement learning among machine learning techniques was employed to develop a semi-active control algorithm for a smart TMD. The smart TMD was composed of magnetorheological damper in this study. For this purpose, an 11-story building structure with a smart TMD was selected to construct a reinforcement learning environment. A time history analysis of the example structure subject to earthquake excitation was conducted in the reinforcement learning procedure. Deep Q-network (DQN) among various reinforcement learning algorithms was used to make a learning agent. The command voltage sent to the MR damper is determined by the action produced by the DQN. Parametric studies on hyper-parameters of DQN were performed by numerical simulations. After appropriate training iteration of the DQN model with proper hyper-parameters, the DQN model for control of seismic responses of the example structure with smart TMD was developed. The developed DQN model can effectively control smart TMD to reduce seismic responses of the example structure.

This research is about a study on the flow stress of Inconel 601 under hot deformation. For Inconel 601, hot compression tests on gleeble 3500 system under 925℃, 1050℃ and 1150℃ and 0.001/s, and 5/s of strain rates were done. The flow behavior of the Inconel 601 was studied and modeled. In this study, the flow stress was modeled using deep neural network and support vector regression algorithm. The flow stress of Inconel 601 was dependent on strain rate and temperature. It was found that both the deep neural network and support vector regression adequately described the flow stress variation of Inconel 601. However, the model by the support vector regression was found to be superior to the model by the deep neural network. The construction of the model by SVR was more efficient than the construction by DNN. Also the prediction accuracy of the model by SVR was better than the accuracy of the model by DNN. It is found that the MAPE(Mean absolute percentage error) of the DNN based model was 4.89% while the MAPE of the SVR based model was 1.98%.

Three CNN (Convolutional Neural Network) models of GoogLeNet, VGGNet, and Alexnet were evaluated to select the best deep learning based image analysis mothod that can detect pavement distresses of pothole, spalling, and punchout on expressway. Education data was obtained using pavement surface images of 11,056km length taken by Gopro camera equipped with an expressway patrol car. Also, deep learning framework of Caffe developed by Berkeley Vision and Learning Center was evaluated to use the three CNN models with other frameworks of Tensorflow developed by Google, and CNTK developed by Microsoft. After determing the optimal CNN model applicable for the distress detection, the analyzed images and corresponding GPS locations, distress sizes (greater than distress length of 150mm), required repair material quantities are trasmitted to local maintenance office using LTE wireless communication system through ICT center in Korea Expressway Corporation. It was found out that the GoogLeNet, AlexNet, and VGG-16 models coupled with the Caffe framework can detect pavement distresses by accuracy of 93%, 86%, and 72%, respectively. In addition to four distress image groups of cracking, spalling, pothole, and punchout, 22 different image groups of lane marking, grooving, patching area, joint, and so on were finally classified to improve the distress detection rate.

Deep learning techniques are being studied and developed throughout the medical, agricultural, aviation, and automotive industries. It can be applied to construction fields such as concrete cracks and welding defects. One of the best performing techniques of deep running is CNN technique. In this study, we analyzed the classification of handwritten images using CNN technique before applying them to construction field. Deep running is generally more accurate with deeper layers, but analysis cost is high. In addition, many variations can occur depending on training options. Therefore, this study performed a parametric study to be a reference when CNN technique was applied through accuracy analysis according to training options.

Global warming due to the increase of greenhouse gases may significantly affect various aspects of the Earth’s environment and human life. In particular, the impacts of climate change on agriculture would be severe, leading to damages to crop yields. This paper examines the experimental prediction of rice yield in China using DNN (deep neural network) and climate model data for the period between 1979 and 2009. The DNN model built through the process of hyperparameter optimization can mitigate an overfitting problem and cope with outlier cases. Our model showed approximately 38.7% improved accuracy than the MLR (multiple linear regression) model, in terms of correlation coefficient with the yield statistics. We found that the diurnal temperature range and potential evapotranspiration were the critical factors for rice yield prediction. Our DNN model was also robust to extreme conditions such as drought in 2006 and 2007 in China, which showed its applicability to the future simulation of crop yields under climate change.

The video game Tetris is one of most popular game and it is well known that its game rule can be modelled as MDP (Markov Decision Process). This paper presents a DQN (Deep Q-Network) based game agent for Tetris game. To this end, the state is defined as the captured image of the Tetris game board and the reward is designed as a function of cleared lines by the game agent. The action is defined as left, right, rotate, drop, and their finite number of combinations. In addition to this, PER (Prioritized Experience Replay) is employed in order to enhance learning performance. To train the network more than 500000 episodes are used. The game agent employs the trained network to make a decision. The performance of the developed algorithm is validated via not only simulation but also real Tetris robot agent which is made of a camera, two Arduinos, 4 servo motors, and artificial fingers by 3D printing.

This paper presents a 6-DOF relocalization using a 3D laser scanner and a monocular camera. A relocalization problem in robotics is to estimate pose of sensor when a robot revisits the area. A deep convolutional neural network (CNN) is designed to regress 6-DOF sensor pose and trained using both RGB image and 3D point cloud information in end-to-end manner. We generate the new input that consists of RGB and range information. After training step, the relocalization system results in the pose of the sensor corresponding to each input when a new input is received. However, most of cases, mobile robot navigation system has successive sensor measurements. In order to improve the localization performance, the output of CNN is used for measurements of the particle filter that smooth the trajectory. We evaluate our relocalization method on real world datasets using a mobile robot platform.

Planetary global localization is necessary for long-range rover missions in which communication with command center operator is throttled due to the long distance. There has been number of researches that address this problem by exploiting and matching rover surroundings with global digital elevation maps (DEM). Using conventional methods for matching, however, is challenging due to artifacts in both DEM rendered images, and/or rover 2D images caused by DEM low resolution, rover image illumination variations and small terrain features. In this work, we use train CNN discriminator to match rover 2D image with DEM rendered images using conditional Generative Adversarial Network architecture (cGAN). We then use this discriminator to search an uncertainty bound given by visual odometry (VO) error bound to estimate rover optimal location and orientation. We demonstrate our network capability to learn to translate rover image into DEM simulated image and match them using Devon Island dataset. The experimental results show that our proposed approach achieves ~74% mean average precision.

In this paper, we propose a jellyfish distribution recognition and monitoring system using a UAV (unmanned aerial vehicle). The UAV was designed to satisfy the requirements for flight in ocean environment. The target jellyfish, Aurelia aurita, is recognized through convolutional neural network and its distribution is calculated. The modified deep neural network architecture has been developed to have reliable recognition accuracy and fast operation speed. Recognition speed is about 400 times faster than GoogLeNet by using a lightweight network architecture. We also introduce the method for selecting candidates to be used as inputs to the proposed network. The recognition accuracy of the jellyfish is improved by removing the probability value of the meaningless class among the probability vectors of the evaluated input image and re-evaluating it by normalization. The jellyfish distribution is calculated based on the unit jellyfish image recognized. The distribution level is defined by using the novelty concept of the distribution map buffer.