|CUI YUNDUAN||1461206: D, 中間発表||知能システム制御 杉本 謙二, 小笠原 司, 松原 崇充, 南 裕樹|
Title: Efficient Reinforcement Learning with Smooth Policy Update in High-dimensional Robot Systems
Abstract: Due to both the limited number of training samples and the curse of dimensionality, it is challenging to apply value-function based Reinforcement Learning (RL) algorithms in real robot systems with a large number of joints. Previously proposed RL algorithms require impractically high computational complexity in time and space for the high dimensionality of the system, and result in unstable learning because of insufficiency of the number of samples. This research focuses on designing efficient algorithms with capability of learning complex tasks in high-dimensional robots. To avoid these problems, we focus on the smoothness of the policy update in RL; If the control policy is updated smoothly, the iteratively generated samples will smoothly cover the state-action space and overcome the instability caused by insufficient and sparse samples. In particular, we focus on Dynamic Policy Programming (DPP) that augments its value function with Kullback-Leibler divergence to control the smoothness of policy update. In this research, we propose to extend DPP by combining it with Nearest Neighbor Search and Kernel Ridge Regression in order to be efficiently updated in high dimensional state-action space. Two algorithms, LUDPP and KDPP are proposed and successfully applied in pneumatic artificial muscle driven robot hand control. The experimental results of both simulation and real robot indicate that our proposed algorithms considerably reduce the computational complexity while have better learning performances compared with conventional ones.
language of the presentation: English
|村瀬 真基||1551106: M, 1回目発表||知能システム制御 杉本 謙二|
title: Study of rubber band handling by dual-arm robot
abstract:Manipulation of deformable objects is still challenging even for state-of-the-art robots. In this study, we focus on handling of a rubber band by a dual arm robot. We discuss an approach to reduce the deformabillity of the rubber band by exploiting its properties. Moreover, some preliminary experimental results are shown.
language of the presentat:Japanese
|LIU CHENGBO||1551203: M, 1回目発表||ネットワークシステム学 岡田 実|
Title: Flexible modulation for OFDM system using parallel combinatory coding
Abstract: OFDM system is widely utilized for many applications including ISDB-T system, 4G and 5G wireless communication systems. However, for some high mobile environment, high received power cannot be guaranteed due to users’ mobility. In addition, some small devices need to use a small size of power amplifier with a limited dynamic range and with a low modulation constellation for application. Therefore, it would be better if we can design a flexible OFDM modulation with low-order QAM or ASK. In this research, we propose two systems which are called half-symbol parallel combinatory multicarrier modulation (PC/HS-MCM) and orthogonal parallel combinatory amplitude shift keying (PC-ASK) modulated multicarrier system, respectively. The former proposed system can achieve better bit error rate (BER) performance and better peak-to-average power ratio (PAPR) performance compared to that of ordinary OFDM system. To further reduce the complexity, we propose the latter system which has a little PAPR degradation compared to that of the first one but still better than that of ordinary OFDM system. The proposed systems can achieve flexible bandwidth efficiency (BWE) range using low modulation constellation which can provide the design freedom degree for system design or adaptive modulation and code control with better PAPR characteristic.
Language of the presentation: English