| 小澤 裕斗 | 1651028: M, 2回目発表 | 知能システム制御 杉本 謙二, 小笠原 司, 松原 崇充, 小蔵 正輝, 小林 泰介 |
|
title:Sparse non-parametric policy improvement with natural image input
abstract:In this research, we propose a reinforcement learning method by pseudo-input Gaussian process regression with natural image input. Simulation and actual machine experiment confirmed the shortening of computation time required for prediction and generalization of learning. language of the presentation:Japanese 発表題目:自然画像を入力としたスパース・ノンパラメトリック方策改善 発表概要:本研究では、自然画像を入力とし、疑似入力を用いたガウス過程回帰による強化学習手法を提案する。シミュレーションと実機実験により、予測にかかる計算時間の短縮と学習の汎化性を確認した。 | ||
| 則永 悠 | 1651088: M, 2回目発表 | 知能システム制御 杉本 謙二, 小笠原 司, 松原 崇充, 小蔵 正輝, 小林 泰介 |
|
title:Policy transfer learning from simulator to real world
abstract: Recent reinforcement learning has been applied with raw image or high dimensional image features as state representation. It has been successful in simulation, however, it cannot be directly applied to real world due to the sample efficiency. One may consider to transfer the learned policy in simulation to real world, however, it is no easy due to appearance errors between the simulation and real world To overcome this limitation, this paper introduces a novel framework of robust feature extraction for ap earance errors between simulation and real world. The proposed method is then applied for policy search in simulation to make the policy transferable to real world. language of the presentation: Japanese 発表題目:シミュレーションから実環境への方策転移学習 発表概要:複雑な環境下で作業するロボットにおいて, センサ情報 から行動への写像を直接学習するEnd-to-end なアプローチが注目されている．しかし, シミュレータと実環境 では見かけ上の違い（アピアランス誤差）があるため, シミュレータ上で学習されたend-to-end な制御方策を直接実ロボットに転移することは難しい．そこで本研究では、Transfer Component Analysis と呼ばれる転移学習を用いて, シミュレータと実環境のアピアランス誤差に頑健な特徴抽出を試みる．そして抽出された特徴量を用いて,シミュレータ上で方策探索を行い，学習された方策が，実環境に転移可能であるかを検証した結果を報告する． | ||
| HAN HAIFENG | 1661024: 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, 小蔵 正輝, 小林 泰介 |
|
title: Model-based Reinforcement Learning Approach for Deformable Object Manipulation abstract: Deformable Object(DO) manipulation has wide application in industry and in daily life. Conventionally, it is difficult for a robot to manipulate a DO to achieve the target configuration due to the absence of the universal model that specifies the DO regardless of the material and environment. Since the state variable of a DO can be very high dimensional, identifying such a model may require a huge number of samples. Thus, model-based planning of DO manipulation would be impractical and unreasonable. In this research, we focus on the subset of DO, called deformable linear object(DLO), and explore the approach based on reinforcement learning. To this end, our approach is to apply a sample- efficient model-based reinforcement learning method to resolve the high dimensional planning problem of DLO manipulation with a reasonable number of samples. To investigate the effectiveness of our approach, we developed an experimental setup with a dual-arm industrial robot and multiple sensors. Then, we conducted experiments to show that our approach is efficient by performing a DLO manipulation task. language of the presentation: English | ||
| 尼子 琢朗 | 1651002: M, 2回目発表 | ロボティクス 小笠原 司, 杉本 謙二, 高松 淳, 丁 明 |
|
title: Construction of non-contact grasping motion measurement system
abstract: Techniques for recognizing the human hand motion are expected to be applied to various fields. For example, designing robot and artificial hands, manipulating machines with gesture interfaces, and developing learning support systems using hands. In this research, To collect data that can be diverted to the grasping strategy of the robot, we construct the system for measuring the human grasping motion. Our method measure the grasping posture, position and force with non-contact with the hand. language of the presentation: Japanese | ||
| 吉田 拓弥 | 1651124: M, 2回目発表 | 大規模システム管理 笠原 正治, 中島 康彦, 笹部 昌弘, 川原 純 |
|
title:Virtual Network Reliability Evaluation by Binary Decision Diagram
abstract: A network reliability evaluation is to find the probability that two specified nodes can communicate with each other in a given network whose links break down with some probabilities. Calculation methods based on Binary Decision Diagram (BDD) are known as strict methods for the network reliability evaluation. Binary Decision Diagram is a data structure that can efficiently express Boolean functions. In this presentation, we propose a method to calculate the reliability of virtual network. Our method does not increase much computation time even if we take link dependencies into account. We conduct numerical experiments, show the results of them and discuss future work. language of the presentation:Japanese 発表題目:二分決定グラフを用いた仮想ネットワーク信頼性評価 発表概要: ネットワーク信頼性評価とは，グラフ中に静的な故障確率が設定されている場合に2頂点間が通信可能である確率を求める問題である． この厳密計算手法として，二分決定グラフ(BDD)による計算方法が知られている． 二分決定グラフは論理関数を効率よく表現できるデータ構造である． 本発表では，仮想ネットワークの信頼性を計算する方法を提案する． 本手法では，仮想ネットワーク上の頂点間に存在する依存関係を二分決定グラフで表現し， 物理ネットワークを表現するBDDとの二項演算により仮想ネットワークの信頼性評価を行う． 本手法を既存手法と比較し，どの程度計算時間が増大するかを述べる． また今後の課題として，この計算時間の高速化手法についても述べる． | ||
| 萬代 光治 | 1651101: M, 2回目発表 | 大規模システム管理 笠原 正治, 飯田 元, 笹部 昌弘, 川原 純 |
|
title: Analysis of cloud platform workload data
abstract: Since demand for cloud platforms such as Google Cloud Platform and Azure is increasing, it is important to improve scheduling efficiency of jobs. However, on large-scale platforms, various kinds of jobs and hardware make job/task scheduling complicated. In order to promote research on the nature and challenges of scheduling, Google released the trace data of a platform actually operated, called Google Cluster Usage Trace. In this presentation, we first analyze the execution history of jobs and tasks included in Google Cluster Usage Trace, and show the trend of execution situation and duration of input job. Next, we describe the result of verifying the waste of computational resources that can be reduced when tasks constituting a job are interrupted based on execution status. language of the presentation: Japanese | ||
| 大神 勝也 | 1651022: M, 2回目発表 | ソフトウェア工学 松本 健一, 安本 慶一, 石尾 隆, 畑 秀明, Raula G. Kula |
|
title: Development of a Real-Time 3D Profiler
abstract: For developers concerned with a performance drop or improvement in their software, a profiler allows a developer to quickly search and identify bottlenecks and leaks that consume much execution time. Non real-time profilers analyze the history of already executed stack traces, while a real-time profiler outputs the results concurrently with the execution of software, so users can know the results instantaneously. However, a realtime profiler risks providing overly large and complex outputs, which is difficult for developers to quickly analyze. In this work, we visualize the performance data from a real-time profiler. We visualize program execution as a three-dimensional (3D) city, representing the structure of the program as artifacts in a city (i.e., classes and packages expressed as buildings and districts) and their program executions expressed as the fluctuating height of artifacts. Through two case studies and using a prototype of our proposed visualization, we demonstrate how our visualization can easily identify performance issues such as a memory leak and compare performance changes between versions of a program. You can see demonstration and download our prototype at https://github.com/sefield/high-rising-city-artifact. language of the presentation: Japanese 発表題目: リアルタイム3Dプロファイラの開発 発表概要: 次のリンクから、デモムービーの視聴および開発した可視化ツールのプロトタイプのダウンロードが可能です。（https://github.com/sefield/high-rising-city-artifact） | ||
| 中才 恵太朗 | 1651080: M, 2回目発表 | ソフトウェア工学 松本 健一, 安本 慶一, 石尾 隆, 畑 秀明, Raula G. Kula |
|
title: The Effects of Donation Badges of Bug Tracking System in the Eclipse Project
abstract: The Eclipse project is collecting donations. Since 2014 November the project has prepared badges for donors on the bug tracking system as one of the benefits. However, the effects of the donation badges have not been clarified. We study the effects of donation badges by analyzing quantitative effects on activities for bug reports using causal inference. language of the presentation: Japanese | ||