21
1 Chapter 12. Chapter 12. Web Information Integration Web Information Integration Using Multiple Character Agents Using Multiple Character Agents Soft computing Laboratory Yonsei University October 27, 2004

1 Chapter 12. Web Information Integration Using Multiple Character Agents Soft computing Laboratory Yonsei University October 27, 2004

Embed Size (px)

Citation preview

Page 1: 1 Chapter 12. Web Information Integration Using Multiple Character Agents Soft computing Laboratory Yonsei University October 27, 2004

1

Chapter 12.Chapter 12.Web Information Integration Using Web Information Integration Using

Multiple Character AgentsMultiple Character Agents

Soft computing LaboratoryYonsei University

October 27, 2004

Page 2: 1 Chapter 12. Web Information Integration Using Multiple Character Agents Soft computing Laboratory Yonsei University October 27, 2004

2

Outline Outline • Introduction• Information integration on multiple character interface• Application prototypes based on the MCI

– Venus and Mars– Recommendation battlers

• Implementation issues of the MCI• An initial evaluation of the MCI using the wizard of Oz

method– Wizard of Oz method– Experiments – Results

• Related work• Conclusion

Page 3: 1 Chapter 12. Web Information Integration Using Multiple Character Agents Soft computing Laboratory Yonsei University October 27, 2004

3

Introduction Introduction • Life-like agent or character (LLA or LLC)

– Software agent with a virtual face and body on a computer display and behaviors like a creature or a person

– Work as an interface between a human user and a computer system

– User-friendly than conventional GUIs

– Advantage To provide an active interface to a system cf. conventional man0machine interfaces

• Web information retrieval – LLA can be applied to help

User-friendly interfaces are welcome Help navigate users to their preferred web pages

• This paper– Discusses a team of agents that work together as mediators

between a user and multiple information sites

– cf. most LLA used work as a standalone guide

Page 4: 1 Chapter 12. Web Information Integration Using Multiple Character Agents Soft computing Laboratory Yonsei University October 27, 2004

4

IInformation integration on multiple character nformation integration on multiple character interfaceinterface

• Information on the web– Tends to be scattered among a number of sites

• Information integration– Scheme to integrate distributed information sites into an

interoperable system– It makes a collection of information sites more valuable than

the individual components

• Conventional information integration system– Designers determine how to integrate the information sites is

specified– User did not know about it– User did not be allowed to change the combination of

information sites nor the integration mechanism

Page 5: 1 Chapter 12. Web Information Integration Using Multiple Character Agents Soft computing Laboratory Yonsei University October 27, 2004

5

MMultiple character interface (1/2)ultiple character interface (1/2)• Multiple character interface (MCI)

– Motivation of MCIEach individual user has different demands or preferences

for information integrationThe best framework is one that allows the user to easily

construct a team of his or her favorite information sites that work together and to customize them flexibly

– Provides an environment where multiple information agents and a human user mutually interact

• Information agent– Body part

Acts as an information gathering engine

– Header part Implemented as an animated LLA

Information integration on MCI

Page 6: 1 Chapter 12. Web Information Integration Using Multiple Character Agents Soft computing Laboratory Yonsei University October 27, 2004

6

MMultiple character interface (2/2)ultiple character interface (2/2)• Communication between user and agent

– User can access the agent by sending a message– Agent can respond to the message by talking with gestures

Information integration on MCI

Page 7: 1 Chapter 12. Web Information Integration Using Multiple Character Agents Soft computing Laboratory Yonsei University October 27, 2004

7

MCI-based agentMCI-based agent• Have some advantages

– Provide a friendly interface between the user and the information sources

– Agents collaborate to assist the user in retrieving and integrating information

– User can easily understand the functionality and role of each information agent by visualizing information agents as characters

Information integration on MCI

Page 8: 1 Chapter 12. Web Information Integration Using Multiple Character Agents Soft computing Laboratory Yonsei University October 27, 2004

8

AApplication prototypes based on the MCIpplication prototypes based on the MCI• Venus and Mars

– Cooperative search engine– Three LLA cooperate with each other to assist an user in

locating cooking recipe pages

• Recommendation Battlers– Competitive restaurant recommendation system– Two LLA compete with each other to recommend restaurants

to a user

Page 9: 1 Chapter 12. Web Information Integration Using Multiple Character Agents Soft computing Laboratory Yonsei University October 27, 2004

9

VVenus and Mars (1/2)enus and Mars (1/2)• Search engine

– Most widely used tools to retrieve information from the web– Not always very useful for novice users such as elderly

people

• Authors utilize domain specific information agents– Provides noiseless information concerning a particular

domain such as recipes, restaurants, or retailers

• Venus and Mars– System that allows information integration based on

keyword associations through conversations among LLCs– Search results are shown in two frames

Left : a list of recipe pagesRight : web page of a list entry when the entry is clicked

Application Prototypes Based on the MCI

Page 10: 1 Chapter 12. Web Information Integration Using Multiple Character Agents Soft computing Laboratory Yonsei University October 27, 2004

10

Venus and Mars (2/2)Venus and Mars (2/2)• Three information agents

– Kon-san

– Cho-san

– Pekko

• Collaborate with each other– Assists in reducing the

number of search results in dialogue steps

– Asks for a tip on seasoning and answers on behalf of the user in utterance step

• Have potential of realizing various types of information search by adding agents to the team

Application Prototypes Based on the MCI

Page 11: 1 Chapter 12. Web Information Integration Using Multiple Character Agents Soft computing Laboratory Yonsei University October 27, 2004

11

Recommendation Battlers (1/2)Recommendation Battlers (1/2)• Electronic commerce (EC)

– One of the most successful application domains of the internet

– Most conventional shopping sites are running in an independent and closed manner

– Comparison shopping sites are run by a third party, which is independent from buyers and sellers

• Recommendation Battlers– New multiagent-based system for EC where multiple

shopping sites or information recommendation sites are integrated in a flexible and interactive manner

– Provides a virtual space where multiple animated agents– Customer compares items recommended by multiple agents

and finds a preferred one by watching a competition performed by the agents on a browser

– Agents can learn his or her preference and use it for further recommendations through interactions with the customer

Application Prototypes Based on the MCI

Page 12: 1 Chapter 12. Web Information Integration Using Multiple Character Agents Soft computing Laboratory Yonsei University October 27, 2004

12

Recommendation Battlers (2/2)Recommendation Battlers (2/2)• Two restaurant recommendation

agents– Peddy– Genie

• Peedy and Genie start to recommend restaurants in a competitive manner after gathering restaurant information from web sites

• Recommendation– Performed by two character

agents interacting with each other and user

– Show the web page that contains the restaurant information

– Add comments about the average cost and the distance from the nearest station

Application Prototypes Based on the MCI

Page 13: 1 Chapter 12. Web Information Integration Using Multiple Character Agents Soft computing Laboratory Yonsei University October 27, 2004

13

IImplementation issues of the MCImplementation issues of the MCI• Architecture of the MCI

– Each agent recognizes actions taken by the user or other agents through data captured by its sensor, interprets the actions, and responds through its actuatorWhen the agent hears something, variables $utterance and

$agent are instantiated

– By combining commands, an agent can perform complicated actions

Page 14: 1 Chapter 12. Web Information Integration Using Multiple Character Agents Soft computing Laboratory Yonsei University October 27, 2004

14

AAgent scenariogent scenario• Agent behavior

– controlled by scenarios written in Q

• Agent scenario– Represented as a state transition graph

Self-introduction Idling

"May I help you?"

Web InformationRetrival

Qustion aboutingredients

Question aboutseasoning

Clicked

Unknownkeyword

The number ofresults is over

1000

Knownkeyword

Knownkeyword

The number ofresults is over

1000

Unknownkeyword

Knownkeyword

Implementation Issues of the MCI

Page 15: 1 Chapter 12. Web Information Integration Using Multiple Character Agents Soft computing Laboratory Yonsei University October 27, 2004

15

IImplementation of MCImplementation of MCI• MCI implement

– Using a control frame and multiple agent frames– When MCI is initialized

Managers are loaded into the control frame

Agent Manager(Java Script)

User Manager(Java Script)

Dialogue Manager(Java Script)

CommandTransmitter

(Applet)

Command Receiver(Applet)

CommandTransmitter

(Applet)

CommandTransmitter

(Applet)

CharacterController

(Java Script)

CharacterController

(Java Script)

...

Control Frame

Agent Frame 1 Agent Frame 2

Implementation Issues of the MCI

Page 16: 1 Chapter 12. Web Information Integration Using Multiple Character Agents Soft computing Laboratory Yonsei University October 27, 2004

16

Wizard of Oz methodWizard of Oz method• Evaluation of Venus and Mars or Recommendation Battlers

system is difficult– They are still at prototype stage– They are not able to communicate with a human user fluently

• Wizard of Oz method– Method to observe the behavior of human subjects toward a

computer system in which a human operator called wizard simulates the whole or a part of the system

– In the paper, the authors modified the Venus and Mars system so the user interacts with wizards through characters

Evaluation of the MCI Using the Wizard of Oz Method

Page 17: 1 Chapter 12. Web Information Integration Using Multiple Character Agents Soft computing Laboratory Yonsei University October 27, 2004

17

ExperimentsExperiments• Three features of MCI

– Multiple characters appear– Characters interact with each other– Characters are heterogeneous and each one has its own role

• Five interfaces used

Evaluation of the MCI Using the Wizard of Oz Method

Number of Characters

Cooperation Roles

A. Cooperative 3 Yes Yes

B. Single 1 - -

C. Chat 0 - -

D. Non-Cooperative 3 No Yes

E. Homogeneous 3 Tes No

Page 18: 1 Chapter 12. Web Information Integration Using Multiple Character Agents Soft computing Laboratory Yonsei University October 27, 2004

18

EExperimental results (1/2)xperimental results (1/2)

Evaluation of the MCI Using the Wizard of Oz Method

Chat vs. Single

Topic Chat Single t-value (n=8) p-value

Specialty 0.2 1.4 -1.26 0.12

Recipe 14.8 13.4 0.88 0.20

Health 0.4 1 -0.8 0.22

Character 0 1.8 -1.61 0.07

Others 4.6 2.4 2.17 0.03

Single vs. Cooperative

Topic Chat Single t-value (n=8) p-value

Specialty 1.4 2.8 -1.18 0.13

Recipe 13.4 10 2.42 0.02

Health 1 2.4 -1.72 0.06

Character 1.8 1.6 0.12 0.45

Others 2.4 3.2 -0.70 0.25

Chat vs. Cooperative

Topic Chat Single t-value (n=8) p-value

Specialty 0.2 2.8 -3.41 0.004

Recipe 14.8 10 3.63 0.003

Health 0.4 2.4 -3.08 0.007

Character 0 1.6 -1.37 0.10

Others 4.6 3.2 1.10 0.15

Page 19: 1 Chapter 12. Web Information Integration Using Multiple Character Agents Soft computing Laboratory Yonsei University October 27, 2004

19

EExperimental results (2/2)xperimental results (2/2)

Evaluation of the MCI Using the Wizard of Oz Method

Non-Cooperative vs. Cooperative

Topic Non-Cooperative Cooperative t-value (n=8) p-value

Specialty 0.2 1.4 -1.26 0.12

Recipe 14.8 13.4 0.88 0.20

Health 0.4 1 -0.8 0.22

Character 0 1.8 -1.61 0.07

Others 4.6 2.4 2.17 0.03

Homogeneous vs. Heterogeneous

Topic Homogeneous Heterogeneous t-value (n=8) p-value

Specialty 1.4 2.8 -1.18 0.13

Recipe 13.4 10 2.42 0.02

Health 1 2.4 -1.72 0.06

Character 1.8 1.6 0.12 0.45

Others 2.4 3.2 -0.70 0.25

Page 20: 1 Chapter 12. Web Information Integration Using Multiple Character Agents Soft computing Laboratory Yonsei University October 27, 2004

20

Related workRelated work• Meta-search engines integrate the output of multiple search

engines and succeed in offering improved performance• In conventional collaborative information integration systems

– Techniques used to coordinate the information agents or information resources are specified by the system designers

– Remain hidden from users

• Andre and Rist propose a system employing multiple characters– Their work mainly emphasizes the advantage of multiple

characters as presentation media

Proposed system in this paper is more like a multiagent system because the information agents are physically distributed over the internet

Page 21: 1 Chapter 12. Web Information Integration Using Multiple Character Agents Soft computing Laboratory Yonsei University October 27, 2004

21

ConclusionConclusion• This paper

– Propose an information integration platform called MCI– Show two application prototypes

Venus and MarsRecommendation Battlers

– Evaluate the MCI by using the wizard Oz method

• Future works– Capability for life-likeness– Capability for collaboration– Capability for presentation– Capability for conversation