TAAI 2016 Keynote Talk: Intercultural Collaboration as a Multi‐Agent System

Intercultural Collaboration as a Multi‐Agent System

Toru IshidaDepartment of Social Informatics

Kyoto University

Part 1: Multi‐Agent Systems: Real‐World ApplicationsPart 2: Intercultural Collaboration: A New  Research Field

Part 3: Multi‐Agent System Problems in Intercultural Collaboration

Part 1

Multiagent Systems: Real‐World Applications

Multi‐Agent Systems

• The research area is often called Autonomous Agent and Multi‐Agent System.

• It models a human society as a collection of independent decision making agents.

• It analyzes and visualizes behaviors of agents via Multi‐Agent Based Simulation.

Multiagent System: A Major Research Area of AI

• IJCAI 2016 （ Int. Joint Conf. on Artificial Intelligence）• Machine Learning: 22 sessions• Agent‐based and Multi‐agent Systems: 12• Knowledge Representation, Reasoning, and Logic: 10 • Planning and Scheduling: 10 

• Textbook of Artificial Intelligence Overview of the bookThe main unifying theme is the idea of an intelligent agent. We define AI as the study of agents that receive percepts from the environment and perform actions……

DAIWS 1980 Launched by Researchers in the United StatesICMAS 1995 International Conference integrating workshops

in the United States, Europe, and JapanPRIMA 1998 Asia Pacific International ConferenceAAMAS 2002 A+ Ranked International Conference

MACC（JP） MAAMAW（EU）

DAIWS（US）

ICMASPRIMA(Workshop)

AAMAS

AA（US）

ATAL（EU）

PRIMA(Conference)

Asia International

1980

2000

1990

2010

How We Developed Our Research Community

Multi‐Agent Systems are Getting Real

Theory（Research for Computational Models of Human Society ）

PRACTICE（Solving Problems for Society）

１９８０ １９９０ ２０００ ２０１０ ２０２０

Multi‐Agent Based Simulation （Economy, Disaster Management）

Matching Algorithm （Medical Care）

Collaboration Distributed Search  Self‐Organization       Multi‐Agent Based Simulation

High Speed ComputingBig Data

Negotiation  Market Auction      Game Theory  Matching Algorithm

Game Theory （Security）

Target #1 Target #2

Target #1 4, ‐3 ‐1,  1Target #2 ‐5,  5 2, ‐1

Example1: Security GameMilind TambeUniversity of Southern California How to allocate limited security resources

• Defenders commit first, and then Attackers observe defenders’ behavior and make decisions.

• Optimal allocation: Weighted randomStackelberg Equilibrium

Contribution to Security

29/52

23/52

22/52 30/52

Utility=0.433

Utility=0.294

Deployed at Los Angeles International Airport

Attacker

Defender 1000 Flights, 20 air marshals          1041

combinations

Example2: Kidney Exchange

• First Idea 1986 [Rapaport] • Exchange Started in 2003 [Roth, Sönmez, Ünver, …]• United Network for Organ Sharing (UNOS) adopted 

Sandholm’s algorithm in 2008 and executes it twice a week.

Tuomas SandholmCarnegie Mellon UniversityIn the US, more than 70,000 patients are waiting 

for kidney transplants, but only 10,000 transplants are carried out each year. 

30‐chain! [New York Times 2/18/2012]

Wife

Husband Mother

Daughter

Contribution to Medical Care 

Price

Share of Tokyo Stock Exchange 

Tick Size 

• If Tick Size (minimum unit of ordering price) is too large, stock prices become unstable.

• A newly discovered threshold made stock prices stable.

• From July 2014, Tokyo Stock Exchange has changed tick sizes of TOPIX100 stocks. This impacted 80% of stock exchanges in Japan.

Example 3: Designing Market Systems by Multi‐Agent Based Simulations Kiyoshi IZUMI

University of Tokyo

Contribution to Economy 

1000 agents with different strategies

Example 4: Decision‐Making for Tsunami Evacuation by Multi‐Agent Based Simulations

• Background • 5,781 people are living with risk of tsunami by earthquake.

• Maintain escape routes to places over 2 meters above sea level.

• Result• Determine bridges for reinforcement.

• Determine the priority of streets to remove snow.

10

Itsuki Noda National Institute of Advanced Industrial Science and Technology

Contribution to Disaster Management

①

② ③

④⑤

⑦⑥

⑧⑨

⑩⑪

Select bridges for  reinforcement from 11 bridges

Condition 1：Prioritize 11 streets to remove snow

→ 211 × 211 ＝ 4,194,304 cases

Condition 2：

Part 2

Intercultural Collaboration: A New Research Field

Intercultural Collaboration

• After 9.11 we created a concept of intercultural collaboration, where participants in the same problem field with different cultures and languages work together towards the shared goal. 

• Because the research target emphasizes collaboration rather than communication, we can clearly identify research objectives; goal‐directed group activities can be evaluated both qualitatively and quantitatively. 

2001

To develop open source software by using machine translation.

China Team

Korea Team

ICE2002 OrganizersTranslation Checker

Japan Team

Malaysia Team

Intercultural Collaboration Experiment (ICE 2002)

Translation among Japanese, Chinese, Korean, Malay, and English

31,000 messages in one year.

Message Input

Translation Button

Translated Massage

Chinese Japanese Korean Malay English

Multi‐Language BBS2002

Translation Pentagon for ICE2002

Arcnet/sangenjaya

Japanese

English Malay

Chinese K orean

J ‐ server

Hard to collect translation engines to cover five languages.Hard to understand their contracts.Hard to evaluate their services.Hard to customize provided services.

2002

The Language GridFragmentation and recombination is the key to create customized multi‐language environments for various intercultural activities.

more more

Disaster Management Education Medical Care

Sharing Multilingual Information Universal Playground Multilingual Reception

Language Support forMulti-language and Multi-cultural Communities

Sharing language resources such as dictionaries and machine translators around the world

German Research Center for Artificial Intelligence

Stuttgart UniversityNational Research Council, Italy Chinese Academy of

Sciences NTT Research LabsAsian Disaster Reduction Center

Kookmin UniversityPrinceton University

NECTECUniversity of Indonesia

Google. Inc.Kyoto UniversityNICT

2007

Dictionary (Data)

Parallel Text (Data)

Human Interpreter

Machine Translation

Wrapping

Wrapping

Wrapping

Wrapping

DictionaryDictionary Service

Parallel TextParallel Text Service

MachineTranslator

Machine Translation Service

Human Translator

Human Translation Service

Accurate wordtranslation

Approximatetranslation

Low qualityhigh speedtranslation

High qualitylow speedtranslation

避難場所

disaster shelter

避難場所は、家から近い学校です。

In the case of disaster, people should be evacuated to a school nearby their house.

避難場所は、家から近い学校です。

Disaster shelter is school close from a house.

避難場所は、家から近い学校です。

Your disaster shelter is the school closest to your house.

From Language Resources to Language Services

Language Grid Operation Centers

Kyoto Operation Center(Kyoto Univ.，2007 ‐ )

Jakarta Operation Center(Univ. of Indonesia, 2012 ‐ )

Xinjiang Operation Center (Xinjiang Univ., 2014‐ )

East Asian Languages

South‐East/South Asian Languages

Central Asian Languages

Share 225 language services from 170 groups in 22 countries

Bangkok Operation Center（NECTEC, 2010 ‐ ）

NSF LAAPS Grid ProjectLDC

EU Meta Share ProjectELRA/ELDA

Federated Operation of Language Grid Servers 

EducationAgricultureKnowledge

Parents

Youth

JapaneseExperts

VietnameseFarmers

JAVI EN

The goal is to increase rice productivity and to decrease the environmental burdens caused by excess use of agrichemicals.

The challenge is to transfer rice harvesting knowledge at realtimefrom Japanese experts to Vietnamese farmers in rural areas with low literacy rate.

A youth‐mediated communication (YMC) model was invented.Children receive IT training, and act as mediators between parents and experts. 

Agricultural Support in Vietnam 2011

Overview of the Project• Organizations

– Kyoto University: Multilingual Communication– NPO Pangaea: Education, Activities– University of Tokyo, Mie University : Agricultural Knowledge– Vietnam National University：Local Arrangements– MARD, DARD: Planning and Controlling Experiments

• Schedule– 2011/02 ~ 2011/03 1st Experiment (Thien My, Vinlong Province)– 2012/10 ~ 2013/01 2nd Experiment (Thien My, Vinlong Province)– 2013/09 ~ 2014/01 3rd Experiment (Thien My  and Dong Thanh, Vinlong Province)– 2014/02 ~ 2011/06 4th Experiment (Dong Thanh, Vinlong Province)

Observation Recording Question/Answer

Example of Experiment Schedule

2012/09 2012/10 2012/11 2012/12 2013/01

Participants: 15 Youth

SowingRice plantingRice reaping

RiceGrowing

YMCExp.

PreparingExperimentWorkshop

2012/10/29 2012/12/152012/12/12012/10/19

Designing Language Communication

Japanese: たんぼの準備として、田起こしや代掻き、あぜぬりをして下さい。Vietnamese: Chuẩn bị đất là kết hợp giữa canh tác đất, cày bừa và đắp bờ.(English: Land preparation is a combination of tillage of the soil, puddling and levee painting.)

Example-based Translation Service

Translation Service with Agricultural Dictionary

Japanese-English

Translator

English-VietnameseTranslator

Multilingual Dictionary

for Agriculture ☆

JapanesePart-of-Speech

Tagger

VietnamesePart-of-Speech

Tagger

Multilingual Parallel Texts for Agriculture

☆☆

Example-Based

MachineTranslator

JapaneseDependency

Parser

BackTranslation

BestTranslationSelection

OnlineConsultationby Human

VietnameseFarmers

JapaneseExperts

☆ Multilingual Dictionary for Agriculture is provided by NPO Pangaea, Japan National Agriculture Research Center, Vietnam MARD. Entry Number: 3,099 (Sep. 2014)☆☆ Multilingual Parallel Texts for Agriculture is provided by NPO Pangaea, Japan National Agriculture Research Center, Vietnam MARD. Entry Number: 2,485 (Sep. 2014)

Best Translation Selection Service

Designing Knowledge Communication

EducationYouthVietnamese

FarmersJAEN

Leaf Color Plate Recipe Card(textbook) 

Measure

Insect Plate

PassportCamera

Parallel Text(compiled)

JapaneseExpertsVI

Parents

AgricultureKnowledge

By NPO Pangaea

Language services are not enough for knowledge communication.

Part 3

Multi‐Agent System Problems in Intercultural Collaboration

Self‐initiated repair or user adaptation to machine translation:  Before posting a message, the sender modifies input texts to improve translation quality. 

Other‐initiated repair or collaborative translation:  If a translation was imperfect, collaboration is initiated by a receiver. 

Topic 1: Collaboration ProtocolInteraction patterns among users in ICE2002.

S. Nomura, T. Ishida, N. Yamashita, M. Yasuoka and K. Funakoshi. Open source software development with your mother language: Intercultural Collaboration Experiment 2002. International Conference on Human‐Computer Interaction, Vol. 4, pp. 1163‐1167, 2003.

2002

Collaborative Translation Protocol

MachineTranslator

Source Language Side

Target Language 

Side

1. Send the source sentence 2. Evaluate fluency of the translation.

3. Send the sentence which is modified to be fluent.

4. Evaluate adequacy of the translation against the source sentence. 

Improve adequacy of the translation result Improve fluency of the translation result

D. Morita and T. Ishida. Collaborative translation by monolinguals with machine translators. International Conference on Intelligent User Interfaces, pp. 361‐366, 2009.

2009

The Vietnam project aims at increasing the yield of rice. One project member, an agricultural expert, recognized the lack of nitrogen in this 

particular rice field, and suggested farmers to use fertilizer a little more. The farmers did not follow this advice, since they believe  bugs will gather from 

neighboring fields, if they increase the amount of fertilizer.

Topic 2: Shaping a Problem2015

Wicked Problem!Wicked problem is difficult or impossible to solve because of incomplete, contradictory, and changing requirements that are often difficult to recognize.

Because of complex interdependencies, the effort to solve one aspect of a wicked problem may reveal or create other problems.

Wikipedia

How to shape a problem?

Topic 3: Problem Solving Organization

JapaneseExperts

VietnameseYouths

VietnameseFarmers

NPOPangaea

Univ. ofTokyo

KyotoUniv.

MARD

MieUniv.

Knowledge Communication

Organizational Communication

JapaneseBridger

EnglishBridger

Language Communication

JAVI EN

VietnameseBridger

VNU

DARD

Agricultural Data

Agricultural Knowledge

Machine Translation

2015

Multi‐Agent Based Participatory Simulation 

Analysis Design

Field Experimentor

Multi-Agent Based Participatory Simulation

Language / Knowledge Communication

ModelingMachine Translatio

n(JA→EN)

Post-Editing

(EN→EN)

JapaneseBridger

VietnameseBridger

Vietnamese Youth

Japanese Expert

Machine Translation

(EN→VI)

Parallel Texts（JA→VI）

Pre-Editing(JA→JA

)

EnglishBridger

Post-Editing(VI→VI)

2015

Test technologies in advance with various stakeholders.

Problem Solving Organization

Project Team for Tien My Commune, Tra On District, Vinh Long Province, Vietnam, Jan 5th, 2013

Problem solving organization guarantees contentious commitment of stakeholders to the problem domain.

To be continued….

• Multi‐agent system research now contributes to our human society.

• Intercultural collaboration can be seen as multi‐agent system research, and will reduce conflicts in the world.

Thank You!

Special Thanks to:

Tuomas Sandholm,Milind Tambe, Bo An, Itsuki Noda, Kiyoshi Izumi, Shigeo Matsubara, Yohei Mukakami, Donghui Lin, Yumoko Mori, Toshiyuki Takasaki