Digital Library[ Search Result ]
Variational Recurrent Neural Networks with Relational Memory Core Architectures
Geon-Hyeong Kim, Seokin Seo, Shinhyung Kim, Kee-Eung Kim
http://doi.org/10.5626/JOK.2020.47.2.189
Recurrent neural networks are designed to model sequential data and learn generative models for sequential data. Therefore, VRNNs (variational recurrent neural networks), which incorporate the elements of VAE (variational autoencoder) into RNN (recurrent neural network), represent complex data distribution. Meanwhile, the relationship between inputs in each sequence has been attributed to RMC (relational memory core), which introduces self-attention-based memory architecture into RNN memory cell. In this paper, we propose a VRMC (variational relation memory core) model to introduce a relational memory core architecture into VRNN. Further, by investigating the music data generated, we showed that VRMC was better than in previous studies and more effective for modeling sequential data.
Cross-Entropy Planning with Prior Updates
HyeongJoo Hwang, Youngsoo Jang, Jaeyoung Park, Kee-Eung Kim
http://doi.org/10.5626/JOK.2020.47.1.88
This paper introduces a method of cross-entropy planning which updates prior probability for planning optimization process. Cross-entropy planning is a popular method in online planning and involves the extraction of samples from a simulation environment and selection of optimal action based on the values of the extracted samples. The performance of the cross-entropy planning is limited due to involvement of optimization processes without usage of previous planning results. We propose a method that updates prior probabilities for the optimization process based on the action sequences acquired from the cross-entropy planning. The proposed method improves the performance of cross-entropy planning with progression of planning epoch. We evaluated the proposed method based on the comparison with the cross-entropy planning in a physical-based simulation (OpenAI Gym) environment.
Analysis of Reward Functions in Deep Reinforcement Learning for Continuous State Space Control
http://doi.org/10.5626/JOK.2020.47.1.78
Deep Reinforcement Learning (DRL), which uses deep neural networks for the approximation of the value function and the policy, in continuous state-space control tasks has recently shown promising results. However, the use of deep neural networks as function approximators has often resulted in intractable analyses of DRL algorithms mainly due to their non-convexities and thus a lack of theoretical guarantee such as asymptotic global convergence of the learning algorithm. Considering the fact that the reward function in reinforcement learning is one of the key entities that determines the overall characteristics of the learning agents, we focused on a smaller but an important aspect of the analysis, investigating the structure of widely used reward functions in DRL tasks and their possible effects on the learning algorithm. The proposed analysis may facilitate identification of appropriate reward functions in DRL tasks, which has often been conducted via trial and error.
Search
Journal of KIISE
- ISSN : 2383-630X(Print)
- ISSN : 2383-6296(Electronic)
- KCI Accredited Journal
Editorial Office
- Tel. +82-2-588-9240
- Fax. +82-2-521-1352
- E-mail. chwoo@kiise.or.kr