• The Journal of the Institute of Internet, Broadcasting and Communication
• /
• v.22 no.6
• /
• pp.75-81
• /
• 2022

Autonomous vehicles, which can dramatically solve the lack of parking spaces, are making great progress through deep reinforcement learning. Activation functions are used for deep reinforcement learning, and various activation functions have been proposed, but their performance deviations were large depending on the application environment. Therefore, finding the optimal activation function depending on the environment is important for effective learning. This paper analyzes 12 functions mainly used in reinforcement learning to compare and evaluate which activation function is most effective when autonomous vehicles use deep reinforcement learning to learn parking. To this end, a performance evaluation environment was established, and the average reward of each activation function was compared with the success rate, episode length, and vehicle speed. As a result, the highest reward was the case of using GELU, and the ELU was the lowest. The reward difference between the two activation functions was 35.2%.

• Journal of the Korea Computer Graphics Society
• /
• v.27 no.5
• /
• pp.55-61
• /
• 2021

The character motion control based on physics simulation using reinforcement learning continue to being carried out. In order to solve a problem using reinforcement learning, the network structure, hyperparameter, state, action and reward must be properly set according to the problem. In many studies, various combinations of states, action and rewards have been defined and successfully applied to problems. Since there are various combinations in defining state, action and reward, many studies are conducted to analyze the effect of each element to find the optimal combination that improves learning performance. In this work, we analyzed the effect on reinforcement learning performance according to the state representation, which has not been so far. First we defined three coordinate systems: root attached frame, root aligned frame, and projected aligned frame. and then we analyze the effect of state representation by three coordinate systems on reinforcement learning. Second, we analyzed how it affects learning performance when various combinations of joint positions and angles for state.

• The Journal of the Korea institute of electronic communication sciences
• /
• v.18 no.1
• /
• pp.99-106
• /
• 2023

Recently, many studies on crop yield prediction using deep learning technology have been conducted. These algorithms have difficulty constructing a linear map between input data sets and crop prediction results. Furthermore, implementation of these algorithms positively depends on the rate of acquired attributes. Deep reinforcement learning can overcome these limitations. This paper analyzes the performance of DQN, Double DQN and Dueling DQN to improve crop yield prediction. The DQN algorithm retains the overestimation problem. Whereas, Double DQN declines the over-estimations and leads to getting better results. The proposed models achieves these by reducing the falsehood and increasing the prediction exactness.

• The Journal of the Institute of Internet, Broadcasting and Communication
• /
• v.22 no.5
• /
• pp.63-68
• /
• 2022

Autonomous driving technology has recently made great progress with the introduction of deep reinforcement learning. In order to effectively use deep reinforcement learning, it is important to select the appropriate activation function. In the meantime, many activation functions have been presented, but they show different performance depending on the environment to be applied. This paper compares and evaluates the performance of 12 activation functions to see which activation functions are effective when using reinforcement learning to learn autonomous driving on highways. To this end, a performance evaluation method was presented and the average reward value of each activation function was compared. As a result, when using GELU, the highest average reward could be obtained, and SiLU showed the lowest performance. The average reward difference between the two activation functions was 20%.

• Annual Conference on Human and Language Technology
• /
• 2022.10a
• /
• pp.289-294
• /
• 2022

의존구문분석은 문장을 의존관계(의존소-지배소)로 분석하는 구문분석 방법론이다. 현재 사전학습모델을 사용한 전이 학습의 딥러닝이 좋은 성능을 보이며 많이 연구되지만, 데이터셋에 의존적이며 그로 인한 자료부족 문제와 과적합의 문제가 발생한다는 단점이 있다. 본 논문에서는 언어학적 지식에 기반한 강화된 지배소-의존소 제약규칙 에지 알고리즘을 심층학습과 결합한 모델을 제안한다. TTAS 표준 가이드라인 기반 모두의 말뭉치로 평가한 결과, 최대 UAS 96.28, LAS 93.19의 성능을 보였으며, 선행연구 대비 UAS 2.21%, LAS 1.84%의 향상된 결과를 보였다. 또한 적은 데이터셋으로 학습했음에도 8배 많은 데이터셋 학습모델 대비 UAS 0.95%의 향상과 11배 빠른 학습 시간을 보였다. 이를 통해 심층학습과 언어지식의 결합이 딥러닝의 문제점을 해결할 수 있음을 확인하였다.

• Journal of the Korea Computer Graphics Society
• /
• v.27 no.3
• /
• pp.13-23
• /
• 2021

In this paper, we proposed a framework that generates the trajectory of a single rigid body based on its COM configuration and contact pose. Because we use a smaller input dimension than when we use a full body state, we can improve the learning time for reinforcement learning. Even with a 68% reduction in learning time (approximately two hours), the character trained by our network is more robust to external perturbations tolerating an external force of 1500 N which is about 7.5 times larger than the maximum magnitude from a previous approach. For this framework, we use centroidal dynamics to calculate the next configuration of the COM, and use reinforcement learning for obtaining a policy that gives us parameters for controlling the contact positions and forces.

• Proceedings of the Korea Information Processing Society Conference
• /
• 2022.05a
• /
• pp.52-55
• /
• 2022

5G 시대에 스마트 모바일 기기가 기하급수적으로 증가하면서 멀티 액세스 엣지 컴퓨팅(MEC)이 유망한 기술로 부상했다. 낮은 지연시간 안에 계산 집약적인 서비스를 제공하기 위해 MEC 서버로 오프로딩하는 특히, 태스크 도착률과 무선 채널의 상태가 확률적인 MEC 시스템 환경에서의 오프로딩 연구가 주목받고 있다. 본 논문에서는 차량의 전력과 지연시간을 최소화하기 위해 로컬 실행을 위한 연산 자원과 오프로딩을 위한 전송 전력을 할당하는 심층 강화학습 기반의 오프로딩 기법을 제안하였다. Deep Deterministic Policy Gradient (DDPG) 기반 기법과 Deep Q-network (DQN) 기반 기법을 차량의 전력 소비량과 큐잉 지연시간 측면에서 성능을 비교 분석하였다.

• The Journal of the Institute of Internet, Broadcasting and Communication
• /
• v.23 no.4
• /
• pp.9-15
• /
• 2023

Existing simple statistics-based recommendation systems rely solely on students' course enrollment history data, making it difficult to identify classes that match students' preferences. To address this issue, this study proposes a personalized major subject recommendation system based on deep reinforcement learning (DRL). This system gauges the similarity between students based on structured data, such as the student's department, grade level, and course history. Based on this information, it recommends the most suitable major subjects by comprehensively considering information about each available major subject and evaluations of the student's courses. We confirmed that this DRL-based recommendation system provides useful insights for university students while selecting their major subjects, and our simulation results indicate that it outperforms conventional statistics-based recommendation systems by approximately 20%. In light of these results, we propose a new system that offers personalized subject recommendations by incorporating students' course evaluations. This system is expected to assist students significantly in finding major subjects that align with their preferences and academic goals.

• Journal of the Korea Institute of Military Science and Technology
• /
• v.27 no.4
• /
• pp.474-484
• /
• 2024

Reinforcement learning, which are also studied in the field of defense, face the problem of sample efficiency, which requires a large amount of data to train. Transfer learning has been introduced to address this problem, but its effectiveness is sometimes marginal because the model does not effectively leverage prior knowledge. In this study, we propose a stochastic initial state randomization(SISR) method to enable robust knowledge transfer that promote generalized and sufficient knowledge transfer. We developed a simulation environment involving a cooperative robot transportation task. Experimental results show that successful tasks are achieved when SISR is applied, while tasks fail when SISR is not applied. We also analyzed how the amount of state information collected by the agents changes with the application of SISR.

• Proceedings of the Korea Information Processing Society Conference
• /
• 2018.10a
• /
• pp.759-762
• /
• 2018

최근 들어 점차 지능형 서비스 로봇들이 인간의 실생활 속으로 들어옴에 따라, 로봇 스스로 다양한 물체들을 효과적으로 조작할 수 있는 지식을 습득하는 기계 학습 기술들이 매우 주목을 받고 있다. 전통적으로 로봇 행위 학습 분야에는 강화 학습 혹은 심층 강화 학습 기술들이 주로 많이 적용되어 왔으나, 이들은 대부분 물체 조작 작업과 같이 다차원 연속 상태 공간과 행동 공간에서 최적의 행동 정책을 학습하는데 여러가지 한계점을 가지고 있다. 따라서 본 논문에서는 전문가의 데모 데이터를 활용해 보다 효율적으로 물체 조작 행위들을 학습할 수 있는 모방 학습과 강화 학습의 통합 프레임워크를 제안한다. 이 통합 프레임워크는 학습의 효율성을 향상시키기 위해, 기존의 GAIL 학습 체계를 토대로 PPO 기반 강화 학습 단계의 도입, 보상 함수의 확장, 상태 유사도 기반 데모 선택 전략의 채용 등을 새롭게 시도한 것이다. 다양한 성능 비교 실험들을 통해, 본 논문에서 제안한 통합 학습 프레임워크인 PGAIL의 우수성을 확인할 수 있었다.