Nitta Lab.
Research Objectives
research topics

人間の知的活動を支援するシステムを開発するには,知識情報処理やヒューマン・インタフェースなどの基礎技術のほかに,応用分野に関する知識の分析とモデリングが必要です.

新田研究室では,人工知能の基礎,応用研究を中心として,交渉エージェント,エージェントシミュレーション,情報可視化,情報基礎,社会エージェント,地域情報処理などの研究を行なっています.

Research Themes
Followings are abstracts of recent researches.
-伴意関係に基づく弁証法的推論法の形式化とその議論への適用-

伴意関係に基づく弁証法的推論法の形式化とその議論への適用[木藤2009]

Mathematical logic is the science that judges the validity of an argument on the basis of the validity of reasoning. On the other hand, we commonly think that a valid argument is the argument that has trusted conclusion. More precisely, we might regard the argument as valid if and only if it has valid reasoning and true premises. Defeasible argumentation provides for the model for logically evaluating the tenability of premises, and on this basis, it enables to evaluate the validity of the argument in line with our daily meaning. Meanwhile, defeasible argumentation shows promise for the basic mechanisms for communication and computation among agents. This is because it is suited for treating inconsistent, uncertain, imperfect, decentralized, and open sources of information multi-agent environment essentially has. In the present study, we focused on social decision making or consensus building in multi-agent argumentation, and realized following two points. Firstly, we defined collaboration, concession, and compromise from the viewpoint of logic, and proposed the reasoning method based on specialization and generalization defined by a logical implication relation. The proposed method enables to draw non-deductive propositions. Secondly, we showed that agents that have the method resolve conflicts dialogically. In other words, agents resolve conflict by producing an alternative solution or hypothesis that is not obtained at the beginning of argumentation.

Appearance of System

妥協案の発見支援

システム全体図

妥協案の発見支援

-Negotiation Support System with Functions Making Arguments and Finding Compromise-

Negotiation Support System with Functions Making Arguments and Finding Compromise[fujii2009]

Most of negotiation support systems handled only numerical business. However it is important to discuss the claim and finding new compromise plan in negotiations. The purpose of this study is developing a negotiation support system with functions debating and discovering compromise. For supporting argumentation, the system calculates the validity of the claim and market rates. There functions enhance satisfaction and fairness of both sides. For finding compromise, the system shows the combinations of demand of users by graph structure. Users can find compromise in viewing of preference of offers and balance of utility of both.

appearance of system

support for finding compromise

appearance of system

support for finding compromise

-Case based On-line Dialogue Support System-

新たな紛争解決手段として注目される “調停” をオンライン上で実現し,また調停を行う調停者の教育支援も行う,オンライン調停支援システムを開発しています.

オンライン調停支援システム

The Development of An Argument Agent in Mediation Training [Maeda2006]

オンライン模擬調停を利用すれば,参加者がインターネットを介して模擬調停 に参加することができます.しかし,模擬調停を行うには,参加者を集める必要が あるため頻繁に行うことができません.
 本研究では,調停者・両当事者の3人が揃わなくても模擬調停を行える環境を つくるため,当事者の一方または両方をエージェントとし人間と論争を行う論争 エージェントを開発しました.論争エージェントに,背景知識と過去の論争ログを利 用し現在の論争状況に適した発言を行わせるため,発言生成や発言選択の技法を 提案しました.さらに,論争エージェントには(1)議論の収束性・発散性,(2)利己的・協調的,(3)論理的・非論理的の3つの特性を設定しました.
  実験により,論争エージェントの3つの特性を変えることで多様な論争を行え ることを示しました.

論争エージェントの手続き

-On-line Dialogue Support System-

We have developed several dialogue agents which interact with people or other agents through a computer network. Negotiation, mediation (ADR) and sales assistant are typical application fields of dialogue agents. Negotiation is a discussion process between people who have different views to reach a goal by exchanging offers, persuading and compromising. ADR (Alternative Disputaion Resolution) is one of negotiation methods which solves disputation by discussion. Mediation and arbitration are included in ADR. Mediation is different from arbitration in a point that the problems should be solved considering psychological viewpoint. Sales assistant is a task to recognize the user's demands and recommend products which satisfy them.

オンラインの「対話」は相手が見えないだけに,意見が過激になったり,交渉が長引いたりしがちです。 われわれは,調停,相談,商品推薦,交渉など,さまざまな対話のシーンにおいて,対話をスムーズに進行させるエージェントの研究をしています。

Online ADR Agent with a Case Base [Tanaka2004]

ADR (Alternative Disputaion Resolution) is one of negotiation methods which solves disputation by discussion. Mediation and arbitration are included in ADR. Mediation is different from arbitration in a point that the problems should be solved considering psychological viewpoint. 取引上のトラブルの解決方法の1つとして調停があります。調停は双方の意見を聞いて, 双方が納得する解決案を求めるものです。 調停は仲裁と異 なり,論理的に問題を解決するだけではなく,双方が納得するまで話し合いを行うため,感情的な面も考慮する必要があります。 ここでは調停事例の データベースを利用し,調停員に次の発言を選択するための判断材料を提供することで,調停を支援するエージェントを開発しています。

オンライン調停エージェント

Conversation Strategy based on Similarity of Speakers[Ueno2006]

This research focuses on the mediator of mediation who is demanded to make it smoothly apply the closure to various persons concerned , and proposed the technique for advising to the mediator by selecting the conversation strategy that adjusted to the person concerned.

インタラクション時の学習

-Legal Reasoning Systems-

When a lawyer solves legal problems, he doesn't apply legal rules to the cases directly because legal rules are provided in the form of natural language and they are ambiguous and not complete from the view of logical systems.

For example, let's consider following rule,

" A vehicle is prohibited to enter the park."

If someone is injured in the park, does an ambulance enter the park?

To solve this problem, simple deductive reasoning is not sufficient. Considering the objective of the rule, a lawyer will interpret the meaning of the rule and apply its interpreted rule to actual cases. They may refer to old cases or may take the customs, public opinion and policies into account when they interpret the rule.

We have developed a model of legal reasoning combining the rule based reasoning, the analogical reasoning, the case based reasoning, the hypothetical reasoning, nonmonotonic reasoning, and so on.

E-learning system which supports legal education [Tanaka2005]

いよいよスタートしようとしている法科大学院においては,どんな学部の出身者であっても, 3年間という短期間に司法試験の受験レベルの専門家を育てる教育が行われます。 法律の初学者に専門知識を教えるには実戦的な教育が有効であり,中でも模擬裁判は有効な教育方法です。 しかし,時間が拘束されるため,教師の負担も大きく,頻繁に開催することは困難です。 そこで,ネットワークを使った模擬法廷を裁判官エージェントが監督し,そのエージェントを教師がモニタすることで, 教師の負担を軽くするシステムを開発しています。

Effective Retrieval of Judicial Precedents based on Their Structure [Egoshi2005]

判例の電子文書化に伴い,目的判例の検索は専門家にとって重要な役割を果たしています. 従来のキーワード検索ではキーワード選択が難しいため,これに代わる,文書を検索質問とする類似判例検索の実現のために, 文書ベクトル手法,及び,判例の特徴を利用した研究を行っています. 内容別に分けられた判例の構造,及び,法的な専 門用語を利用するために,XMLタグによる争点分割と適合性フィードバックを用い て,より法的な判断部分の類似した判例を検索することを目指しています.

Case Retrieval

Vidualization of Logical Structure of Judicial Precedents [Muramatsu2002]

裁判の結論は判決文としてまとめられます。 判決文の中は,原告の請求内容,事実の概要,争点に関する原告と被告の主張,裁判官の判断などが含まれています。 過去の判例は,非常に重要な情報源なので,判例データベースは法律家にとって必須のものです。 しかしながら,判決文は非専門家には非常に読みにくく,中に展開されている論理を追跡するには時間がかかります。 そこで,判決文の中に〈事実〉,〈理由〉,〈根拠〉などを表すタグを埋め込み,それを利用して, 個々の争点に関する論理の展開を図示するシステムを開発しました。

Disputation Diagram

Research of an agent which supports E-commerce [Nakatsukasa2000]

インターネットの普及に伴い,近年,ネット上でのショッピングが盛んになってきています。 しかしながら,消費者に法律上の知識に乏しいため,取引上の思わぬトラブルに巻き込まれる可能性も増加してきています。 そこで,ネット上の商取引に関して,個々のアクションでどのような法律効果が生じているか, また,トラブル回避のためにどのようなアクションを起こせばよいかを助言するエージェントを開発しました。

E-commerce support agent

法的論証生成支援システムの構築 [加藤1997]

ルールの間に優先関係が存在する場合,ある結論が成立するかどうかを証明するために,その結論の論証を求めるだけではなく, その論証に対する反論,その反論に対する再反論。。。,などを求め,論証の間の優先関係を計算して, その結論の論証が反駁されないことを示すシステムを開発しました。

-Multi Modal User Interface-

"Multi Modal Information Processing" is information processing using various types of information such as image signals and phonetic signals. We are sure that multi modal information processing is key technology for smooth interaction among people.

We have developed a multi modal debating system, Mr. Bengo, with the multi modal processing group of Electrotechnical Laboratory, Professor Ishizuka (Univ. Tokyo) and Associate Professor Okumura (Tokyo Tech.). Mr.Bengo is a demonstration system in which a plaintiff agent and a defendant agent dispute by exchanging facial expression.

Strategy of Facial Expression of Animated Agents [Yuasa2004]

ネットワークを介して交渉を行う場合に,単に条件を相手に提示するだけでなく, 擬人化エージェントの表情を通じて心理状態を相手に伝達した方が交渉が円滑にまとまることがある。 たとえば左の図にあるように,匿名の者同士の価格交渉に表情データを用いることによって, 交渉経過がある程度,コントロールできることが観測されている。 このような表情の戦略は,取引の交渉だけでなく,損害賠償の交渉や,調停や仲裁や,教育の場面での応用が期待できる。

Online Mediation System with a Sense of Presence based on Haptic Interaction [Chen2009]

Haptic interaction is a big project for the next-generation communications. Our research aims at developing an online mediation system which can create a virtual spaces for users to share. Instead of the traditional input devices (such as keyboard and mouse) used for the online mediation system, we utilize a haptic device named PHANToM and the speech recognition feature.With the 3D virtual objects and interface, the users can feel the sense of co-presence. What's more, using the speech recognition feature ,the emotion parameter can also be created and visualized. By the integration of visual and haptic information, the users can feel high realistic sensation during the online conversation.

Online Mediation System with Haptic Devices
Online Mediation System with Haptic Devices
-Fire-Evacuation Simulation-

The Role and Evaluation of Leaders by Fire-Evacuation Simulation[Mitsubayashi2009]

In this study, we assumed a tall building in case of a fire, and some of residents became a temporal leader to lead the evacuee. We conducted fire-evacuation simulation. It turned out that there is a role of leaders that increases the chances of survival and decrease the evacuation time.

Number survivor and evacuation time
Number that evacuate in smoke
Number survivor and evacuation time
Number that evacuate in smoke
-Ad hoc Communication-

Process Estimation of Word-of-mouth Information Spread Based on Ad Hoc Communications[Takei2009]

It can be considered that ad hoc communications is more effective for the collection of local information based on the action. However, there are problems of the unnecessary information flood. In this study, we propose a method of estimating process of Word-of-mouth information spread based on analysis of the human network generated by using a contact history among people and data of infected person with information and the data of infected person with information. By the analysis of a generated network, it becomes possible to estimate the people which transmitted information and the events that led to the information diffusion.

Human network generated by using a contact history among people
Human network generated by using a contact history among people

Agent Community Analysis in Ad Hoc Communications[Makimoto 2005]

It can be considered that the information exchange with the adjacent terminal using ad hoc communications is more effective for the collection of local information based on the action. However, there are problems of the information flood and the failure to transmit and problems of user's receiving unnecessary information according to the forwarding method.

To solve these, we generate the agent who send and receive on mobile devices,and we use the "Agent Community" formed by these agents' communications.

We analyze how the Agent Community is formed by the change of the forwarding condition and the action situation,and we evaluate the effectiveness of the information transmission using the agent community.

Ad hoc Communication Simulator
Ad hoc Communication Simulator
-RoboCup Soccer Simulation-

RoboCup is a landmark project for technology of robotics and Artificial Intelligence. We have participated in Simulation League.

Learning and Adaptation in Dynamic Multiagent Environment [Akiyama2004]

RoboCupとは, ロボット工学と人工知能の融合,発展のために自律移動ロボットによるサッカーを題材として提唱されたランドマークプロジェクトです. 本研究は,その中でも実機を扱わないシミュレーションリーグを対象としています. プレイヤとして動くエージェントは,不完全でノイズが含まれる知覚情報しか得られません. 更に,コンピュータゲームなどとは異なり,全てのエージェントは異なるプロセスで動作するプログラムであるため, プレイヤは完全に独立して制御されます. われわれは,このような複雑な環境の中で学習・適応を行う手法の研究を行っています. 詳細はSoccer Simulationのページへ.

The RoboCup Soccer Simulator 2D The RoboCup Soccer Simulator 3D

RoboCupサッカーシミュレーションにおけるプレイヤエージェントの動的ポジショニングモデル[Hasegawa2007]

エージェント間の仮想的な単方向バネを接続したスプリングモデルにより,味方プレイヤの位置を考慮した動的なポジショニングを可能としている.

The Dynamic Positioning Model

Learning Multiagent Coordinated Behavior in Keepaway Task [Isano2009]

In early works of Keepaway which is the subtask of RoboCup, reinforcement learning was applied only to an agent with a ball for acquiring coordinated behavior. In this study, reinforcement learning was applied to agents without a ball either. In addition, the reward function of the agent with a ball was corrected, and the dribble was added to the action choices. As a result, it came to be able to select a flexible action according to the situation, and a short ratio at the task continuance time decreased at the same average keeping time as the early work.

Keepaway
Getting variable of state
Keepaway
Getting variable of state
Passer dribbling and Receiver moving
Receiver’s reward function
Passer dribbling and Receiver moving
Receiver’s reward function
-RoboCup Rescue Simulation-

Task allocation for heterogeneous agents in RoboCup Rescue Simulation[uehara2006]

The RoboCup Rescue Simulation is a challenge that aims for a robust city for large-scale urban disasters. This system simulates an earthquake, fires and building collapses. In the simulated disaster situation, efficient activities by rescue agents are absolutely essential. The purpose of this research is an efficient task allocation, and focusing on the dependence between heterogeneous agents to achieve it.

Rescue Simulator
Rescue Simulator
Task Allocation
Task Allocation

-Analysis on Change of Research Trends Based on Academic Papers-

Analysis on Change of Research Trends Based on Academic Papers[Okuoka2009]

We developed the system which acquires the tendency of the academic conference and research fields by generating a network based on the prepublication paper. We performed the documents investigation of four representative academic conference, the system can visualize/analyze the different network based on the document data. The system is equipped with the function which is useful for analyses, such as the difference graph and field-related detailed display, search support, statistical information. Moreover, we think that it becomes a useful tool for a researcher if it comes to be provided on the homepage of the academic conference.

User Interface of System
User Interface of System
Field-Related Detailed Display
Difference Graph
Field-Related Detailed Display
Difference Graph
-A Scenario Extraction Support System using KeyGraph and Data Crystallization from Utterance Logs-

We dealt with a negotiation log that was one of utterance logs that the amount is large, and developed a system from which the user was able to understand the outline of the log easily. The system makes topics being discussed and the relations between the topics visible by using the feature of the negotiation log.

A Scenario Extraction Support System using KeyGraph and Data Crystallization from negotiation Logs [Zeze 2008]

In the electronic documents, there are utterance logs. The utterance log is a log that is converted spoken language that two or more participants spoke a text file. It is rare that the point of an argument of the utterance log is consistent because various topics arise in the discussion and progress correlatively. Moreover, reading the log manually and understanding the outline is work to require time and the labor very much because the amount of the utterance log increases if the discussion becomes long.

In this research, the negotiation log with large amount like this was dealing with and we developed a system from which the user was able to understand the outline of the log easily. The topics being discussed and the relations between the topics are made visible by using the feature of the negotiation log. The system extracts the topics by using prior knowledge of the negotiation log and gives the chance for the user to understand the entire structure like the relations between topics from using the technique of Data Crystallization. Moreover, the user can analyze the topics hierarchically by cutting out the range of the analysis because the topics included in the negotiation log were a layered structure.

A Graph by the proposed Scenario Extraction Support System

-交渉ログからのダイアグラム作成による交渉テキストマイニング-

論争構造を定義するダイアグラムを会話のテキストログから作成するシステムを開発し,そのシステムの結果を用いて,交渉の評価の補助を行う研究をしています.

Analysis of Negotiations with Argumentation Diagrams[Miura2007]

近年,実践的な交渉教育として,実際の事例を用いて模擬の調停を行う模擬調停を行い,それを評価することにより参加者の交渉スキルの向上を図る教育が行われている.
しかし,模擬調停の評価は主に論争のプロセスを評価するものだが,評価は1つの論争を見た審判員同士が話し合って決める方式や,論争のログを読んで複数人が良い論争を投票で決める方式のものであった.
そこで本研究では,論争プロセスの評価のために,論点とリンク,そして論点の出現条件となる単語の共起条件からなる論争ダイアグラムを用いて共通のオントロジ作成を支援し,そこで作成したオントロジを用いて定義した4つの観点から評価をするための情報のマイニングを行う手法を提案した.
そのオントロジ作成ツールを用いたオントロジ作成支援の効果と,論争評価における4つの評価基準の妥当性を評価するために実験を行った.その結果,情報の抽出とオントロジ作成に関する提案手法の有効性の確認と評価基準の妥当性を示した.

論争ログからの解析の流れ

-Network of Academic Fields based on Documents Information-

Network Analysis of Academic Fields based on Documents Information[Shimizu2007]

本研究では,ジャーナル・学会ごとの分野分布や分野の流行や廃りを見るための分野キーワードをノードとしたネットワークグラフを作成する手法および視覚化するツールを提案している. 文献情報に基づき,研究分野をノードとしたネットワークグラフを作成し,分野間の関係を視覚化する.本研究では付加情報に基づく ネットワークとテキスト情報に基づくネットワークの2種類の分野間ネットワー クを提案しており,開発した視覚化ツール(図)でこれら2種類のネットワークを 視覚化し,分析する.2種類のネットワークのどちらを用いても学会同士の分野 分布についての比較や流行・廃りを知ることができる.視覚化ツールはグラフ同 士の比較や統計情報の出力などの機能を備えている.電子図書館などに応用できれば研究者にとって役立つ情報を与えることができるようになると考えている.

研究分野ネットワークの視覚化の様子
-Classroom aided system using button device-

Classroom aided system using button device[Kita2009]

高校の授業を対象とし, 特殊な装置を用いず簡易なマウスを利用することで,生徒に授業への参加を促しながら, 生徒の反応を収集し, 集計や一覧表示させて, 授業を支援するシステムを開発した. 実際の授業で使用し,その上で授業の進行を阻害せず, 教師の負荷を増やさないように考慮しつつ, 生徒および教師へのフィードバックも行っている.

データ収集と生徒へのフィードバックによるシステム構成
データ収集と生徒へのフィードバックによるシステム構成
-Automatic Model Selection via Annealing Corrected Error Backpropagation-

Automatic Model Selection via Annealing Corrected Error Backpropagation[Sekino2009]

We investigate a learning method based on the corrected log-likelihood for hierarchical models. Variable selection of linear regression models by calculating AIC of each combination of variables needs exponential order of calculation. We propose "Annealing Corrected Error Backpropagation" which can obtain the solution by an annealing optimization.

ACEBP_BOSTON
Annealing
ACEBP_BOSTON
Annealing
-Learning Algorithm for Hierarchical Models-

In this work, we explain why overfitting occurs when the number of hidden units of the learning model is too much and propose a new learning algorithm for hierarchical models.

Unbiased Likelihood Backpropagation Learning [Sekino et al, 2007]

Unbiased Learning for Hierarchical Models [Sekino et al, 2007]

It is known that overfitting occurs when a conventional statistical learning method is applied to hierarchical models. In this paper, we explain why overfitting occurs when a conventional statistical learning method is applied to hierarchical models, and propose a new learning algorithm for hierarchical models.

Unbiased Learning
-Self Organizing Function Approximation Model-

Self Organizing Function Approximation Model "Self Organizing SiG pseudo kernel Map" [Sekino2005]

There are locally weighted models using the Gaussian Radial Basis Function, RBF network and Normalized Gaussian network (NGnet). NGnet realizes better approximation performance than RBF network by extrapolating generalization based on the normalizing process. In an optimization process of these locally weighted models, parameters such as center position, covariance matrix and estimated value of each basis are the target of optimization, because of too high degree of freedom of these models, we need much sample data for estimation.

In this work, a new locally weighted model which constrains a degree of freedom by setting variance of each basis based on distances to neighbors is proposed. At first, this work proposes an asymmetric basis function "SiG pseudo kernel", which is able to be set extension in arbitrary directions using eigenvalue-eigenvectors. Using this SiG pseudo kernel, the new locally weighted model "SiG pseudo kernel Map" is constructed. SiG pseudo kernel Map connects neighbor relationships of centers of the SiG pseudo kernels, and set eigenvalue-eigenvectors based on the location relationships with the neighbors. No eigenvalue-eigenvectors is set in direction to no neighbors, therefore extrapolating output is done without the normalizing process. In addition, it is possible to limit a degree of freedom of the model by constraining overlap parameter of variance to the neighbors. By means of extrapolating generalization and constraining a degree of freedom, it is expected that SiG pseudo kernel Map realizes high generalization performance.

Because variance of each basis is dependent on location relationships with the neighbors, modification of a center position of a basis affect neighbor's variance. As an optimizing method of SiG pseudo kernel Map, this work also proposes "Self Organizing SiG pseudo kernel Map (SOSiGmap)", which self organizes SiG pseudo kernel Map by means of iterative modification of center position.

Self Organizing SiG pseudo kernel Map
-Grading Support System of Short Essay-

We are developing the grading support system that can reduce human labor by grading examination papers automatically and highlighting the related text.

Grading Support System of Short Essay using SVM [Murata 2007]

Although there have been automatic grading systems of essay, these are not for the examination that required detailed logic, but for the free-stating essay. So, in this research, we have developed the grading support system in order to grade many examination papers effectively and stably.
This system is the grading support system including automatic grading function. In this automatic grading stage, at first, this system extracts linguistic features from the sample papers graded by human-raters. And then, this system predicts the scores of test papers by making SVMs learn these features. The results that the system predicts are displayed on GUI, and then, human-rater gives a final evaluation by confirming and editing these results. In this way, the reliability of grading has been improved by human check after the automatic grading. In this check work, the text and the key words related to each grading item are highlighted on the screen, which supports human confirmation and editing work.
From the results of the experiments, we have confirmed that the automatic grading of this system is as certain as human grading and that this system shortens the time to grade all papers.

Proposed System
-Multi User Learning Agent-

Using the experience of interaction with other users, MULA can adapt to users quickly with similes and differences.

Multi User Learning Agent with Clustering [Ohmura 2007]

We think, in human society, agent need take care multiple user. Then, we propose a learning method for an agent to interact with other agents effectively. This method, MULA, improves efficiency of the learning, and influences the learning experience of one agent to other agent with similarities. Similarity among agents is evaluated by comparing experiences of agents. And, Humans generally classifies massive information into coherent semantic units and accumulates them. This mechanism is called Categorization. We developed MULA-C which is a learning method which allows learning from past similar experiences by using a clustering algorithm.

MULA
-Extraction and Analysis of Adaptive Action Skill of Robot by Multiple Human Direct Teaching-

We have developed a novel machine learning technique by which a robot not only aquires knowledge autonomously but receives advices from a human instructor.

Extraction and Analysis of Adaptive Action Skill of Robot by Multiple Human Direct Teaching[nyuwa2007]

The purpose of this research is to develop a system to obtain the evaluation values and to feed back for teacher, teacher can change the teaching policy. In this system, we obtain the evaluation value of an action of a robot by using robot simulator Webots. Feedback is achieved by gradual clustering from current teaching data and past teaching data, and by presenting the representative instruction of the cluster displayed on the simulator. In the experiment, we confirmed that the representative instruction data in each cluster was improved by clustering with the adding evaluation values, and humans' instruction was improved by referring to the representative instruction data.

Proposed System

Appearance of the experiment

Proposed System

Appearance of the experiment

-Human Agent Interaction by Social Interaction-

To aquire social knowledge for a robot, the robot has to interact with various people. We have developed machine learning technique which aquires social knowledge effectively by using experience with many people.

Action Learning System using Similarity of Interaction with Users [Ohmura2004, Katagami2005]

In general, it is said that social intelligence is not made from the interaction with one person but will arise by the interaction with two or more persons in the social science field. To construct the autonomous system with human-like flexibility, the achievement of such social intelligence becomes an important key. However, so far the research has often designed correspondence with the multi-users or the multi-environments by the top down approach. It is difficult to build agents that have an intrinsic notion of sociality, that develop social skills and bond with people, and that can show flexible correspondence. Then, a bottom-up approach that acquires the society from the interaction with the user of the environment is researched. However, the amount of research of social intelligence using such an interaction is small.

In this research, we aim to achieve a social interaction based on the social learning studied from the interaction with two or more users by the composition theory approach. Social Interaction is realized by two learning functions. One is direct learning function using individualization of user parameters, and another is indirect learning function using past experience and the similarity between each user. In particular, indirect learning uses experience with other users for the target user. We think it is one form of social learning to realize the suitable action from experience with two or more users to the current user.

システム概要図
実験の様子
Overview of the learning system
Snapshot of the experiment

Development of Social Adaptive Agents[Ohmura2009]

Recently, agents are being used actively in various human’s society. We note that human adaptability to society because of making the social agent which is capable of adapting to human society. Therefor we develop social adaptive agents which are MULA and implicits rules learning Agent.

MULA(Multi User Lerarning Agent)
MULA(Multi User Lerarning Agent)
Implicit Rule Leraning Agent
Implicit Rule Leraning Agent
-Interactive Evolutionary Robotics-

We have developed a novel machine learning technique by which a robot not only aquires knowledge autonomously but receives advices from a human instructor.

Automonous action learning system using user's operation instructions [Katagami2003]

Under the situation that it’s difficult to prepare the knowledge for action, the autonomous robot requires the ability to accomplish tasks in the environment where contents human activities. The situation, for example, may be an dynamic environment or unexpected interaction from human. Therefore, the study for acquisition of autonomous behavior and adaptation to various environments become necessary.

Recently, reinforcement learning and evolutionary computation technique were used as the framework of learning and adaptation. Moreover, the research that enables a robot get a controller autonomously has attracted attention. When making interaction dynamics with robot’s embodiment and environment reflect in construction of a controller, one of the purposes of these techniques is eliminating the unsuitable and unnecessary bias by the designer. Therefore, in the former, it has been made usual to learn by trial and error to an agent, without putting in prior knowledge in the framework of reinforcement learning. However, the execution speed becomes a problem in applying to a real environment.

Then, some researches of the approach using interaction with the human who exists in environment has been carried out. Particularly for the robots that do not have a priori knowledge or commit trial and error in the initial stage, human instruction is the very effective acquisition technique of autonomous behavior. However, in a certain level of autonomous robot, it is not necessary to follow instruction from human all the time. In the stage which does not need instruction, robot should demonstrate its autonomy based on the instruction rules stored by interaction with human without putting a burden on human. Therefore, we need to the technique of establishing a robot’s autonomy from through interaction between human and a robot is required.

Our purpose is realizing a robot’s autonomy by receiving the instruction information as a suitable act from human, and gaining act rules evolutionally with the state recognition which can solve a task. We call such a framework Interactive Evolutionary Robotics (IER). In this paper, we propose Active Teaching method taught by timing regarding a teacher’s cognitive load in IER. We compared it with previous teaching methods by simulation experiment.

interface

実験環境

Interface of the System

Experimental Environment

-Information Recommendation System on a Map-

Route Navigation System using the Information Submitted to the Web Page [Yamamoto2004]

We have developed a system which recommends the route and the destination using other user's posted information.

インタフェース

The Credibility of the Posted Information Based on a Map[Yamamoto2005]

We propose a method for estimating the credibility of the posted information from users. Since posted information can include subjective information from various perspective, we can't trust all of the postings as they are. We propose and integrate two factors of credibility of posted information. First, credibility of vote by other users is determined based on the posting user's own credibility and other users' credibility who approved or disapproved her/his posting information. The other is credibility of geographic posting tendency of the users. This credibility is determined based on dispersion of posted information on a map.

-Visualize and Verify System of Contribution Information-

Visualize and Verify System of Contribution Information[Shimizu2006]

This research focused on disaster information contributed to the Internet bulletin board. It has problems because information is locally, the whole image is not seen easily, and the credibility is unknown. Consequently, the choice of information is difficult. We developed the system that has purpose to solve these problems. This has the function to visualize whole image by automatically choosing contribution information using geographic condition and the framework that verifies credibility of uncertain information by using spread forecast by agent simulation.

Screenshot of The System

-Previous Works in Nitta Lab.-

Web-based English Words Learning System[zhu2006]

As it is boring to study English words, it is quite difficult to keep study motivation. In this research, we propose a new concept called WordMap and an English word learning system, which applies WordMap to assist user of English word study, and motivate user's interest for studying English. By using WordMap, English learners can grasp the bias of their vocabulary visually and map out their learning strategy easily.

Word Map
Overview of the learning system
Word Map
Overview of the learning system

A Framework of Comparing Semantics of Nonmonotonic Reasoning [Shibasaki2003]

非単調推論とは,新しい事実が付け加わったときに,以前に成り立っていた知識が成り立たなくなる推論のことです。 今までに,多くの非単調推論の意味論が提案されてきています。また,それらを比較するための枠組みの理論も研究されています。 しかしながら,従来の枠組みでは,知識の間に優先関係がある場合には適用することはできませんでした。 そこで,従来の枠組みを拡張し,優先関係のある場合にも適用できる枠組みを提案しました。

Computing Preferred Answer Sets in Answer Set Programming [Wakaki1998]

知識の間に矛盾があるとき,知識の間に優先順位をつけて矛盾を解消する必要があります。 制約条件を満たしながら,目的の結論を得るためにどのような知識間の優先順位を付けたらよいかを計算する手法を提案しました。

Development of Unified Knowledge Base [Jo2003]

インターネットを介して,さまざまな生命データのデータベースにアクセスすることができます。 しかし,それらのデータベースを利用するには,専門的な知識を必要としますし,データベース間の連携もよくありません。 そこで,生物の知識が十分でなくとも,複数のデータベースを統合的に検索する知識ベースの仕組みを研究しました。

Simulation of Signal Transmission of Genome Signal in RNA [Ikawa2002]

細胞は外部からシグナルを受け取ると,さまざまな化学反応が連鎖反応的に生じ,シグナルが伝達していきます。 その反応の系列を決定するには,まず,反応系列を予測し,実験により,反応の生成物の濃度変化が予測どおりに生じることを確認します。 このようなシグナル伝達モデルの構築を支援するため,データベースを検索して伝達ネットワークを構築する機能と, ネットワークをレイアウトする機能と,化学反応をシミュレーションする機能を持つ,伝達モデル作成支援システムを開発しました。

Modelling Support System for Transmission Model

確率付文脈自由文法によるRNA2次構造予測 [Kawamura2000]

DNAやタンパク質の配列解析に隠れマルコフモデル(HMM)が有効であることは知られていますが, RNAの2次構造の解析するには表現力が十分ではありません。 そこで,確率的文脈自由文法(sCFG)を使って2次構造予測をする手法を開発し,精度の高い予測ができることを示しました。

Prediction of topical subject based on link information on Web [Ohkubo2005]

We have developed techniques to analyze various documents such as Web pages, judicial precedents, technical papers, and so on. These documents are transformed as XML documents and treated uniformly.

The purpose of this research is to predict the subjects which will become the fashion in an electronic bulletin board in the near future. We proposed the technique which analyzes propagation of the subject based on link information. To extract the pattern of propagation, we proposed several criteria to measure the fashion degree of the subject based on link information which appears in contributed articles. We realized prediction method with unknown subject in fashion using the classication by Support Vector Machine. We conducted experiments to verify the validity of this technique with known collected fashion-subjects.

リンク情報の解析

Genaration of Slide Documents from Technical Papers [Takeichi2002]

 われわれは学術論文を書き,それを外部で口頭発表すときには,OHP資料を作成します。  OHP資料を作成するには,論文中から図や表を抽出し,文章を抜き出したり要約したりすることが必要となります。  発表時間に応じて枚数を変えたり,相手に応じて内容を変えたりすることは手間のかかる作業です。  学術論文の多くはLaTeXで清書するため,論文中に文書の論理構造を示すことができるLaTeXのコマンドを使っています。  そこで,このコマンドをベースにして,論文中にその構造を示すタグを付加し,このタグを使って,OHP資料の作成を支援するシステムを研究しました。  下の図は,システムによって与えられた枚数割り当てに応じて,OHP書類の編集をしているところです。

Genaration of slide documents

分析問題の難易度調整支援システム [Ochi2002]

就職試験などで出題される分析問題(与えられた条件から,成立する性質を導く問題)を作成するには, その問題を解くのに要する時間を見積もる必要があります。 この解答時間を調整するため,可能な解の数を計算し,表示するシステムを開発しました。

Dialogue Agent which Recommends Products [Nagai2005]

DVDレコーダのようなサイクルが短く,専門性の高い商品に対しても,精度の高 い推薦を行えるような商品推薦システムを開発しています.あまり知識を持たな い人でも使えるように,対話型エージェントが知識や情報を教示しながら,要望 を明確化していきます.また,電子掲示板上での相談事例を自動抽出し,利用す ることで,情報の移り変わりの速さに対応し,多様な意見を取り入れた推薦を実 現します.

Agent which Recomments Products

Extracting Consultation Rules from Cases for Online ADR [Nishihara2004]

取引上のトラブルがあっても,裁判まで進むことはまれで,各地の消費者センターなどで相談されることも少なくありません。 消費者センターなどでは,相談の内容に応じて,解決案を助言したり,専門家を紹介したり,場合によっては調停や仲裁を行う可能性もあります。 そのような相談員を育てるには時間がかかり,また,近年,相談件数が増大しています。 そこで,典型的な問題に関して相談ルールを自動抽出し,相談員がいなくても典型問題に関して相談を行うWebページの生成システムを開発しました。

Consultation Rules

Monopoly Agents [Oguchi2001]

モノポリーはサイコロを振ってコマを進めるボードゲームの一種です。交渉により,土地の権利書などをうまくそろえた人が大儲けをします。 上級者のゲームでは1つの交渉に数分かかることもめずらしくありません。 上級者のゲームをビデオにとり,それを解析することによって,上級者の交渉戦略を検出しようとしています。 また,抽出した戦略の有効性を検証するため,モノポリーサーバを作成し,異なった戦略を持つプログラム同士でゲームができるようにしています。

事例を用いた秘書業務の支援システム [Suzuki2001]

オンラインで他の人に連絡を取りたい場合であっても,相手が忙しいときや不在のときには連絡がつかないことがあります。 そこで秘書エージェントを使って,伝言を受け付けたり,連絡内容の重要度に応じて取り次ぎを行ったりするシステムを開発しました。

エージェントによる3次元部品構成のモデル生成 [Nakahara2001]

複数の3次元形状のデータにノイズがある場合において,形状の特徴ごとに1つのエージェントを割り当て, エージェント間の交渉によって,形状のクラスタリングをする手法を開発しました。

ゲーム理論を応用した交渉支援システムのエージェント化 [Satoh2000]

交渉において,提案を出すだけでなく,その妥当性の検証を行い,反論や供託や脅しを含めた交渉戦略を提示するシステムを開発しました。

マルチエージェント環境における集団交渉戦略 [Masushio2000]

複数エージェントが派閥を作って交渉を行う場合,派閥の大小によってどのような交渉戦略をとるのが最適か, また,個々のメンバーにとって最適な派閥はどのようなサイズか,を強化学習によるエージェントシミュレーションによって決定します。

交渉支援システムにおける意思決定支援機能について [Yamazaki1999]

法律問題に関する交渉を支援するため,判例の検索,判例の多変量解析,AHPによる意思決定支援などの機能を持つシステムを開発しました。

交渉会話の状態遷移モデルに関する研究 [Kitoh1999]

提案と論証(説明,説得)を含む複雑な交渉の状態遷移モデルを提案し,交渉者の戦略により, 交渉がどのように進展するかをエージェントシミュレーションで解析しました。


Nitta Laboratory
Department of Computational Intelligence and Systems Science
Interdisciplinary Graduate School of Science and Engineering
Tokyo Institute of Technology