KSC 2018
Current Result Document :
| 한글제목(Korean Title) |
Reptile-First-Order Meta-Learning Implementation for Pendulum Reinforcement Learning Problem |
| 영문제목(English Title) |
The space time scan statistics-based analysis of online shopping preferences using parcel shipping data |
| 저자(Author) |
Quang Nguyen
Ngo Anh Vien
TaeChoong Chung
|
| 원문수록처(Citation) |
VOL 45 NO. 02 PP. 0761 ~ 0763 (2018. 12) |
한글내용
(Korean Abstract) |
|
영문내용
(English Abstract) |
Meta Learning has been grabbed much attentions recently after a lot of significant improvements in deep learning. As usual, a very large deep learning neural network is trained from scratch and learns to handle a specific task according to knowledge gained from a very large number of observations presented to the system during the training time. Although the trained model can handle the trained task properly after sufficient training, it hardly generalizes this good performance to other similar but new tasks, especially if constrained with only few-shot learning. Meta learning is introduced as a learning method for the network so that it can be trained in the condition of data sparsity for faster convergence compared to learning from scratch. Introduced in 2018, Reptile, a simple algorithm for First-Order Meta-Learning, has shown its simplicity and robust as an initialization method in dealing with an entire task distribution. To address the challenge of applying Reptile to Reinforcement Learning problems as stated in its original paper, the authors will propose an implementation of Reptile to Pendulum environment with the hope to propagate this algorithm in the Reinforcement Learning area. |
| 키워드(Keyword) |
|
| 파일첨부 |
PDF 다운로드
|
