情報数理学専攻 平成31年度情報数理学セミナー



日時: 12月12日(木) 14:40~16:10
会場: 情報科学研究科A棟 A210・A212会議室
内容: 博士学位論文公聴会
講演者: SOPCHOKE SIRAWIT 14:40〜15:25
講演題目: Explainable and Unexpectable Recommendations using Relational Learning on Multiple Domains (説明可能で意外な推薦のための複数ドメイン関係学習)
概要: In this thesis, we combine relational learning with multiple domains to develop a formal framework for recommendation systems. The design of our framework aims at: (i) constructing general rules for recommendations, (ii) providing suggested items with clear and understandable explanations, and (iii) delivering a broad range of recommendations including novel and unexpected items.
We use relational learning to find all possible relations, including novel relations, to form the general rules for recommendations and explanations. The rules are learned from data and probabilistic setting using Inductive Logic Programming (ILP). The data used is prepared in prolog format. The predicates and all related information can be defined in the setting. The output learned rules will be ranked based on their probabilities and be used to compute the recommendation items. The top-N items in the list will be recommended to a particular user. Our framework aggregates all information and knowledge from multiple domains to broaden the search space, thereby significantly increase a chance to discover items which are previously unseen or unexpected to a user resulting in a wide range of recommendations.
The main contributions of this thesis are as follows: (i) A new recommendation systems framework is developed to construct general rules for recommendations. Each rule is represented in relational logic associating with probability. The rules are used to suggest the items in any domain. (ii) The explanation is inducted from user preferences and all the knowledge available. The explanation describes precisely how the recommendation was selected for a particular user and it (the explanation) is simultaneously formed within the recommendation rule construction process with no additional complicated process required. (iii) Our framework provides suggested items with clear and understandable explanations. It is in if-then logical format which is unambiguous, less redundant and more concise compared to a natural language and other forms (e.g., visual image, tag cloud) used in other explanation recommendation systems. (iv) Our framework delivers the items which are normally unfamiliar to the users or are hidden from them. The recommendation is no longer centered around very small popular items in the head and the long tail problem can possibly be overcome. (v) Lastly, our framework is extendible in that the new domains (e.g., TV show, sport, beauty), new data (e.g., user/item features) and context (e.g., time) can directly be incorporated without the need of re-training. The more knowledge is provided to the framework, the more specific and useful item will be recommended to the particular user.
The experiment results show that our framework is very promising as it does produce interesting recommendations, not found in the primitive single-domain based system, and accompanied with simple and readable explanations.
講演者: 西崎 陽平 15:25〜16:10
講演題目: 機械学習を用いた波面計測・波面制御の高度化
概要: 補償光学は波面計測、波面制御の要素技術から構成され、波面ゆらぎの影響を受ける光学システムの結像精度を高める上で重要な技術である。 その歴史は長く、生体イメージングや天体観測において様々な手法が提案されている。 一般的に波面計測、波面制御はハードウェアの複雑化、複数枚撮影や反復演算が必要である。 一方、近年の情報科学技術の急速な進歩により、深層学習等の機械学習が注目されている。 光計測、光制御においても機械学習の積極導入が進み、光学分野の発展に寄与している。
本論文では、波面計測・波面制御技術に対し機械学習を導入し、設計自由度を拡張させたシングルショット一般化波面計測、ロバスト位相回復および高精度波面制御手法について実証する。 波面計測に対して機械学習を導入することにより、波面センサの設計自由度を大幅に拡大させた。 点光源あるいは2次元拡張光源を用いた簡易な光学変調による波面計測により単一強度画像からシングルショット波面計測を実現させた。 また、位相回復手法における反復型手法と機械学習による非反復型手法について再構成精度、推定速度およびロバスト性の点で数値実験に基づいた比較を行った。 その結果、機械学習による非反復型手法の優位性を示すことができた。 最後に、条件の異なる複数のホログラムを用いて再生波面の精度向上を図った。 単一の可変型ホログラムに比べ、コントラストの向上やアーティファクトの低減の点で再生品質が向上することを数値実験によって明らかにした。 これら成果は、機械学習が波面計測、波面制御に及ぼす効果を数値実験、光学実験を通じて示しており、本研究が既存の光学システムにおける課題を解消する可能性を実証している。


日時: 12月19日(木) 13:00~14:30
会場: 情報科学研究科A棟 A210・A212会議室
内容: 博士学位論文公聴会
講演者: 植松 直哉 13:00〜13:45
講演題目: Efficient Branch-and-Cut Algorithms for Submodular Function Maximization
概要: When we approach problems such as sensor placement, influence spread, and feature selection problems in computer science, we often encounter situations where we need to select a subset of a finite whole set for maximizing some utility functions. Some of such utility functions are known to be submodular. Our study focuses on the submodular function maximization problem under a cardinality constraint which is NP-hard. For the problem, it is known that a variety of greedy algorithms obtain good feasible solutions, quickly. However, some applications seek optimal solutions or solutions better than a solution obtained by a greedy algorithm. Nemhauser and Wolsey (1981) formulated the problem into a binary integer programming (BIP) problem with a huge number of constraints, and proposed a constraint generation algorithm that starts from a reduced BIP problem with a small subset of constraints taken from the constraints. Their algorithm repeats solving a reduced BIP problem while adding a new constraint at each iteration. Unfortunately, their algorithm has to solve too many reduced BIP problems until obtaining an optimal solution. Therefore, we propose an improved constraint generation algorithm which generates a promising set of constraints at each iteration. To improve the efficiency, we incorporate it into a branch-and-cut algorithm. Moreover, real applications such as feature selection (Das and Kempe, 2011) sometimes do not hold exact submodularity. Therefore, as an extension of the submodular function maximization problem, we also consider a maximization problem with a function which is an approximate version of submodular function. We first formulate the problem into a BIP problem with a huge number of constraints and a submodular ratio which represents how much a function is close to submodular. Finally, we propose three exact algorithms including a branch-and-cut algorithm based on the BIP formulation. For the both problems, the computational results for three types of well-known benchmark instances showed that our branch-and-cut algorithms performed better than the conventional exact algorithms.
講演者: 趙 宇 13:45〜14:30
講演題目: Nonparametric estimation of productivity changes using Malmquist-type indices
概要: Productivity growth plays an essential role in both micro- and macro-economics, as it reflects the long-term improvements in production and operations at the firm, industry, and economy-wide levels. There is a wide variety of measures of productivity change, but the Malmquist-type indices are particularly noteworthy because of its widespread use in the literature on productivity. The essential characteristic of Malmquist-type indices is its dynamic view of efficiency, whereas the original efficiency analysis has been mostly static. Therefore, the thesis covers both the theoretical and practical topics of efficiency and productivity analysis to estimate the productivity change with Malmquist-type indices. The primary analyzing approach of this thesis is nonparametric in the sense that the measurement of the production frontier is entirely based on the observed input-output data. The thesis extends the theoretical and practical framework of two principle nonparametric methods involved: Data Envelopment Analysis (DEA) and Stochastic Nonparametric Envelopment of Data (StoNED). Based on the theory of DEA, we developed a new efficiency concept: allocative efficiency regarding profit-ratio maximization. The derived efficiency is then used to construct a novel comprehensive productivity index: a profit-ratio change index. Meanwhile, based on the theory of StoNED, we also proposed a stochastic nonparametric estimation of Malmquist-type indices to account for the impact of noise. The main contributions of this thesis include the following: (1) a new scheme of allocative efficiency, which provides a comprehensive understanding of the sources of inefficiency in inputs and outputs, (2) a new Malmquist-type index termed profit-ratio change index, which gives a full picture of the sources of productivity change in the sense that the impact of allocative efficiency changes are incorporated, (3) a new panel-data model for estimating the Malmquist-type indices under stochastic noise, where we addressed the issues of inconsistent inefficiency and measurement issues of intertemporal inefficiency. Further, the merits of the proposed methods and the validity of the evaluation results have been illustrated in empirical applications. We analyzed the efficiency and productivity change of samples of 37 Japanese securities companies and 101 Japanese regional banks, respectively. We also investigated the drivers of productivity change by applying the decomposition of Malmquist-type indices. These results provide realistic projections and policy implications for improving the productive performance.
Keywords: Productivity growth; Malmquist-type indices; Data Envelopment Analysis (DEA); Stochastic Nonparametric Envelopment of Data (StoNED); nonparametric approach; Japanese banking data set.



日時: 4月18日(木) 14:40~15:40
会場: 大阪大学 吹田キャンパス 共通講義棟 U2-211
内容: 安全教育講演Ⅰ
司会: 齋藤 真人 助教


  1. 14:40~15:10 齋藤 真人 「化学・生物学に関する安全講習」
  2. 15:10~15:40 馬越 貴之 「コンピュータセキュリティ」


日時: 5月9日(木) 13:00~15:40
会場: 情報科学研究科A棟 A109講義室
内容: 博士論文中間発表会
講演者: 康 子辰(鈴木研) 13:00~13:30
講演題目: Analysis of Dynamical Properties of Reservoir Computing using Coupled Time-delay Elements
概要: The reservoir computing scheme is a machine learning mechanism which links the computational capabilities of dynamical systems to information processing. It has been proven both in experiments and theory that even a single-variable time-delay system can act as a reservoir. Although single-variable time-delay systems can efficiently perform information processing, their computational power is limited by the architecture with simplicity. Networks of coupled time-delay elements can provide some useful insights for enhancing speed, performance and memory capacity of the reservoir computing based on time-delay systems. However, the relationship between the dynamical properties of coupled time-delay elements and its computational power is not clearly understood. In this study, we focus on a reservoir of coupled time-delay elements and analyze the system numerically by studying the dynamical properties such as the Lyapunov exponents. The aim of this research is to investigate the relation between the computing performance and the dynamical properties of the reservoir of coupled time-delay elements.
講演者: Emerico Habacon Aguilar(藤崎研) 13:30〜14:00
講演題目: Agent-Based Models for the Opinion Dynamics of Online Social Networks
概要: Online social networking applications play a significant role on how people form their views about various issues that affect social and political landscapes. This motivates us to understand how opinions and influence spread among users of online social networks. Several models exist for describing the opinion formation process among a group of individuals. In this research, we consider agent-based models for modeling the opinion dynamics of online social networks.
These models have mathematical properties that can help in better understanding how communities reach an agreement or disagreement and how fast opinions converge. However, the behavior of these models varies based on the structure of the networks and the type of interactions they represent. We extend the gossip algorithm used in distributed network systems to incorporate group interactions that can represent the communications that take place in online social networks. Using simulations, we show that while this model achieves consensus, the magnitude of change in opinions varies depending on the size of the group that an agent interacts with. We also present an index for describing different variations of the gossip algorithm based how they reach consensus.
講演者: 下村 優(谷田研) 14:00〜14:30
講演題目: DNAマイクロマシンの実現に向けたDNAゲルの光形成・光分解
概要: マイクロマシンによる物質の移動操作技術の実現は,生命現象の解明や細胞組織の人工的形成につながる.マイクロマシンの動作には光制御が有効であり,非接触で遠隔な操作を可能とする.しかし,光制御のためのマイクロ構造をマシンに搭載する必要があり,可動域が制限され自由度の高い動作が困難であった.本研究では.DNAゲルによって構築される光制御型マイクロマシンによる移動操作技術の確立を目指している.DNAゲルは塩基配列の設計や分子修飾により所望の反応特性を有する.光信号に応じてDNAゲルの形成・分解に基づく粘性・形状変化によりマイクロマシンの柔軟な作が可能となり,物質の運動制御が実現される.本報告では,マイクロマシンの開発動向・技術的限界と本研究の意義を明らかにし,これまでの研究成果を述べる.
講演者: Emsawas Taweesak(沼尾研) 14:40~15:10
講演題目: Transfer Sequence Learning in Music-Emotion Recognition using Wearable physiological sensors
概要: Recent studies in brain-computer interface (BCI) have facilitated and stimulated the development of systems and sensors that recognize and interpret human affects. Wearable physiological sensors have been developed to monitor health activity and also utilize knowledge in many research areas. Meanwhile, emotion recognition based on physiological signals has been a hot topic and leverages techniques from multiple areas, such as signal processing and machine learning. For music-emotion recognition, recording physiological responses help to acquire an actual user's feedback while listening to the music.
Unfortunately, in physiological research, there are a small number of benchmark emotional physiological databases and it consumes a long time to acquire them. Nevertheless, machine learner needs a lot of training data to perform pattern-learning ability. To mitigate this shortcoming, an emerging technique called transfer learning is adopted. The main idea is overcoming the isolated learning paradigm and utilizing the knowledge acquired for one task to solve related ones. A database for emotion analysis using physiological signals (DEAP) dataset, a multimodal dataset for the analysis of human affective states, is used to train a convolutional neural network (CNN) to represent learned features.
For sequence-to-sequence learning, the recognition techniques learn to interpret sequence between music and emotion periodically. The main idea can divide into window recognition and sequence learning. Window recognition can recognize emotion by using a sliding window and then accumulate the result respectively. Considering window learning, the classifier can only learn the information from each window without knowledge in the previous period. Hence, this research presents a new type of sequence learning that recognizes emotional affects and memorizes previous information throughout learning. Long short-term memory (LSTM) network is an efficient sequence learner with the ability to preserve information overtime by memory cells and gate units. Therefore, sequence learning outperforms window recognition in continuous emotion recognition.
This study proposes transfer sequence learning in music-emotion recognition using wearable physiological sensors. The experiment was designed to show the performance of sequence learning and transfer learning and to compare our method with traditional techniques. We investigated emotion based on electroencephalogram (EEG) and also used the DEAP's EEG dataset to train feature extractor by using CNN to avoid data limitation. Besides, LSTM network was used to recognize sequence data and achieve higher accuracy than traditional techniques.
講演者: Bassel Ali Ashour(沼尾研) 15:10~15:40
講演題目: Reinforcement learning based Distance Metric Filtering Approach in Clustering
概要: Conventional feature selection methods may not provide the sufficient means to deal with the diverse growing amount of data nowadays. Evolutionary Distance Metric Learning (EDML) relies on an evolutionary approach in its distance metric learning process; this process in case of diagonal EDML can be viewed as an embedded feature weighting one. However, such process is done simultaneously on all features and does not explicitly select the features. This research introduces a hybrid system called R-EDML, combining the sequential decision making of Reinforcement Learning (RL) with the evolutionary feature prioritizing process of EDML in clustering. The goal is to create a feature selection control strategy that aims to optimize the input space by reducing the number of selected features while maintaining the clustering performance. This can lead to future data collection time and cost reduction. In the proposed method, features represented by the diagonal elements of EDML distance transformation matrices are prioritized by a differential evolution algorithm. Then a selection control strategy using Reinforcement Learning is learned by sequentially inserting and evaluating the prioritized elements. This process is repeated with the aim to optimize the matrices by filtering the elements used in them. The outcome is the selection of the best R-EDML generation matrices with the least number of elements possible. Diagonal matrix R-EDML is compared to normal EDML and to conventional feature selection with EDML in terms of feature selection and accuracy. Results show a decrease in the number of features compared to EDML and to conventional feature selection with EDML, while maintaining a similar or higher accuracy level. Full matrix R-EDML which focuses on diagonal and non diagonal elements is tested and shows good results as well. Finally, R-EDML policy which chooses the features is tested for each EDML generation and tested unified across all generations and both ways show promising results and potential for this research. Future work aims to tackle the dimensionality problem of EDML by using higher dimensional data sets and different RL techniques and approaches to handle them. Furthermore, we aim to increase the accuracy by modifying the EA using RL as well as testing new changes to EDML and use new synthetic data sets along with real data sets in the experiments.


日時: 5月16日(木) 14:40~15:40
会場: 大阪大学 吹田キャンパス 共同講義棟 U2-211
内容: 安全教育講演Ⅱ


  1. 14:40~15:10 小西  毅 「電気、電子機器、電磁波、赤外・紫外光、レーザー光」
  2. 15:10~15:40 木村 吉秀 「各種機械・工具・工作機械、放射線、粒子線」


日時: 6月20日(木) 14:40~16:10
会場: 大阪大学 吹田キャンパス 情報科学研究科A棟 A109
内容: 講演会
講演者: 中川 正基 特任助教
講演題目: オンオフ間欠性の統計的性質と数理モデル―確率過程からカオス,そして無限峰写像―
概要: オンオフ間欠性とは、静かな状態と激しく変化する状態が不規則に現れる現象のことである。 様々な物理系や実験系で観測されており、その統計的性質を理解するための簡単な数理モデルがいくつか存在している(相乗確率過程や結合カオス系、変調ロジスティック写像など)。 それらの数理モデルの多くには共通した構造があり、共通した統計的性質が成り立つことがわかっている(オンオフ間欠性の「標準統計法則」)。 しかしながら、いくつかの現象では、標準統計法則では捉えられない性質が観測されており、標準を超えた数理モデルの開拓が必要と考えられている。 本講演では、オンオフ間欠性の統計的性質を理解するための様々な数理モデルを紹介し、そこから導かれる統計法則について説明する。 さらに、講演者の研究の中から、標準を超えた統計的性質をもつ"新奇な"数理モデル(無限峰写像によるオンオフ間欠性)を紹介する。


日時: 6月27日(木) 13:30~15:40
会場: 大阪大学 吹田キャンパス 情報科学研究科A棟 A109
内容: 講演会
講演者: Elisa Capello 博士 (イタリア トリノ工科大学 / CNR-IEIIT) (13:30~14:30)
講演題目: Flight Control System Design and Identification for a Multirorotor UAV
概要: The purpose of this seminar is the introduction of the students to the practical approach for the design of the control laws implemented in the onboard Flight Control System (FCS) of multi-rotor Unmanned Aerial Vehicles (UAVs). In the last decade, different indoor flight navigation systems for small UAVs have been investigated, with a special focus on different configurations and on sensor technologies. The main idea of this research is to propose a distributed Guidance Navigation and Control (GNC) system architecture, in which both a flight controller and a companion computer are considered. A method for the identification of the quadrotor parameters, in terms of provided thrust and torque is also included. The key features of this application are: (i) robustness of the proposed control methods, (ii) data fusion of different sensors and (iii) robust and flexible configuration. Both theoretical and practical parts of FCS are analyzed, focusing on indoor applications. From classical to variable structure controllers are proposed. Simulations and flight tests are performed with some of these controllers.
講演者: Elisabetta Punta 博士 (イタリア CNR-IEIIT) (14:40~15:40)
講演題目: Sliding Mode Control: a Survey of Strategies and Applications
概要: The control of dynamical systems in the presence of uncertainties is a common problem to deal with when considering real plants. Actually, the real plant behaviour is affected by uncertainties the influence of which should be carefully taken into account when considering the system performance. For this reason, the control of uncertain processes has attracted great interest in the research community (Corless and Leitmann 1981, Doyle et al. 1994, Ho and Khalil 1997, Young et al. 1999, Serrani et al. 2001). Among existing methodologies, the sliding mode control (SMC) technique (Slotine and Li 1991, Utkin 1992, 1999) turns out to be characterized by high simplicity and robustness. The main idea at the basis of SMC techniques is that of designing a sliding surface to which the controlled system trajectories must belong. On the sliding manifold the behaviour of the system is the expected one and is insensitive to model uncertainties and disturbances. The real-life implementation of SMC techniques presents a major drawback, due to the finite switching frequency of real control devices. The high-frequency components of the control could excite parasitic resonant modes so that the system trajectories largely differ from the ideal ones. The system and the actuators non-ideal behaviour can produce the so-called chattering phenomenon, which is a high frequency motion that makes the state trajectories rapidly oscillating about the sliding manifold. Chattering and the need for discontinuous control constitute two of the main drawbacks of Variable Structure Systems (VSS) with sliding modes. The most straightforward approach proposed in literature to avoid chattering is to approximate the sign function of the discontinuous control by the saturation function. As a result, the system motion is confined within a boundary layer of the sliding manifold (Slotine and Sastry 1983, Burton and Zinober 1986, Slotine and Li 1991). Nevertheless, if the parasitic dynamics is not well modelled and taken into account, the approximation of the discontinuous control could compromise the disturbance rejection properties of SMC (Young et al. 1999). A different solution is to embed an asymptotic state observer into the controller so that the discontinuous control is confined within a high-frequency loop by-passing the real plant (Young and Kwatny 1982, Utkin 1992). A different approach to avoid chattering is to augment the controlled system dynamics, by adding integrators at the input channel, so as to obtain a higher-order system in which the actual control signal and its derivatives explicitly appear. If the discontinuous signal coincides with the highest derivative of the actual plant control, the latter results continuous with a smoothness degree depending on the considered derivative order. This procedure refers to higher order SM (Emelyanov et al. 1986, Levant 2001), dynamic SM (Sira-Ramirez 1993), and terminal SM (Man Zhihong et al. 1994). The higher order SM research line led to control algorithms belonging to the family of 2-SM controllers, that is algorithms in which the relative degree between the constraint output (zeroing which the system motion meets the desired performance specification) and the discontinuous control is two. When dealing with this kind of controllers for uncertain systems it is necessary to solve differential inequalities of order greater than one. Most of the existing results in nonlinear control theory (including Lyapunov stability theorems) rely on the comparison principle, which in general does not work in the high relative degree case. Special treatments, often non-systematic, are needed. In this seminar various SMC algorithms and strategies will be presented in a unified framework. Finally, some recently presented applications of this class of algorithms will be surveyed.


日時: 7月4日(木) 14:40~16:10
会場: 大阪大学 産業科学研究所 講堂
内容: 講演会
講演者: 木村 司 助教
講演題目: 心理学における実験研究とその実例
概要: 製品の品質向上やサービスの多様化に伴い、商品そのものの価値(モノ)ではなくそれに付随する体験や感覚の価値(コト)が重要視され始めている。 これらの付加価値はユーザの心理状態と関連しており、多くの分野でこれらのデータの取得や分析が試みられてる。 その中で、心理データをどのように取得すればよいか、また、どのような場面でどのようなデータを取得することが有用であるか、など信頼性の高い心理データを取得する方法論が求められている。 本講演では心理学における心理データ取得のための研究法を概説し、主観、行動、生理に代表される心理データの有用性や適用範囲を説明する。 さらに、これらの方法論に基づき講演者が実施した心理実験について紹介する。


日時: 7月11日(木) 13:00~14:30
会場: 大阪大学 吹田キャンパス 情報科学研究科A棟 A109
内容: 講演会
講演者: 水谷 康弘 准教授 (工学研究科 機械工学専攻)
講演題目: 弱値が切り開く新たな光計測 ~量子光学からAIまで~
概要: 一見、繋がりがなさそうに見える量子光学と情報工学は、近年のデバイスの進展と解析技術の発達により弱い相互作用を計測するという観点において、その概念が繋がりを持とうとしている。 本講演では、それら2つの学問領域の中間領域ともいえる機械工学の視点から両者を俯瞰し応用展開している計測方法について紹介する。 例えば、光スピンホール効果とよばれる反射光の位置がnmオーダーでシフトする量子光学的な現象を測定することでサブナノオーダーの表面粗さをとらえている。 一方で、擾乱環境下におけるわずかな揺らぎの違いをとらえることで、100光子数以下のイメージングも光相関イメージングも可能になった。 さらに、AI技術を導入することで劇的に高速化を実現した手法についても紹介する。



日時: 7月25日(木) 14:40~16:10
会場: 大阪大学 吹田キャンパス 情報科学研究科A棟 A109
内容: 講演会
講演者: 杉江 俊治 先生(京都大学名誉教授/大阪大学・コマツみらい建機協働研究 所・特任研究員)
講演題目: カーネル法のシステム同定と制御への応用
概要: カーネル法は機械学習の分野で良く知られた手法であるが、近年ではシステム同定の分野でも注目を集めている。 比較的短い入出力データから対象システムのインパルス応答を高精度に求めること事ができることが一つの特徴である。 本講演では、第一に、カーネル法の利点について、数値例を交えて直感的に説明した後、これをシステム同定に有効に適用する手法について紹介する。 第二に、この同定実験に際しての、入力の選択法について述べる。 最後に、カーネル法が、実験データに基づくPID制御のオートチューニング等の制御手法にも有効に働くことを、実験及び数値例により示す。


日時: 10月10日(木) 13:00~17:00
会場: 情報科学研究科A棟2階会議室 A210&A212
内容: 修士論文中間発表会


日時: 10月17日(木) 13:00~14:30
会場: 大阪大学 吹田キャンパス 情報科学研究科A棟 A109
内容: 講演会
講演者: 梶野 洸(日本アイ・ビー・エム株式会社 東京基礎研究所)
講演題目: 分子ハイパーグラフ文法とその分子最適化問題への応用
概要: 分子最適化問題とは所望の物性値を持つような分子を発見する問題である。 この問題に取り組むためには主に2つの課題を解決する必要がある。 1つめの課題は、化学的に妥当な分子を制御可能な形で生成することである。 分子の生成にあたっては、例えば原子価などの陽に書ける制約を守る必要がある。 2つめの課題は、物性値の計測やシミュレーションのコストが高いことが多いことである。 そのため、なるべく少ない計測回数でより良い分子を見つける必要がある。 多くの既存研究は、分子生成に特化した変分オートエンコーダ(VAE)とベイズ的最適化の組み合わせにより上記の2つの課題に取り組んでいるが、原子価の制約を100%満 たせない手法が多く、100%満たせるものであってもニューラルネットワークの構造が複雑で訓練が難しい。
本研究では、より簡単な構造のVAEを用いてこの問題を解決することを目指す。 特に制約を守った分子を必ず生成できるグラフ文法(分子ハイパーグラフ文法)を補助的に用いることで、標準的なseq2seq-VAEでこの問題を解決できることを示す。 本研究の主な貢献は、分子ハイパーグラフ文法の提案およびその文法をデータから構築するアルゴリズムの提案である。 実験を通じて(1)計測回数制約の有無にかかわらずVAEベースの手法の中で最も良い性能であること(2)計測回数制約があるもとでは強化学習ベースの手法よりも良い性能であることを実証する。


日時: 11月14日(木) 13:00~13:30
会場: 情報科学研究科 A棟 A109講義室
内容: インターンシップ報告会


日時: 11月28日(木) 14:40~16:10
会場: 大阪大学 吹田キャンパス 情報科学研究科A棟 A109
内容: 講演会
講演者: 松井 知己(東京工業大学工学院経営工学系・教授)
講演題目: スポーツスケジューリング
概要: 従来、スポーツ競技のスケジュール(試合日程)は手作業で立てられることが多かったが、 コンピュータの支援の下にスケジュールを作成する手法の研究が、 数年前よりアメリカを中心に大きく発展している。 メジャースポーツにおける試合日程は、観客数や試合中継等による収入にも影響を与えることから、スポーツビジネスに おける重要な問題として注目を浴びつつある。 日本においても、Jリーグの試合日程作成は数年前よりソフトウェアを用いて行われている。 本発表では、総当たり戦のスケジュールの作成法について議論する。 総当たり戦のスケジュールの作成はグラフ理論や組合せ理論の分野において古くから議論されている問題であり、 その性質について様々な議論がなされている。 ここで議論するスケジュールは、日程だけでなく試合場の割当等も同時に考慮したものとなっている。 本発表では、スポーツのスケジュール作成に用いられる、数学的な性質や最適化手法について説明し、 スケジュール作成 法に関する近年の結果を紹介する。