1

Dynamic Generation of Agent Communities from Distributed Production and Content-Driven

Delivery of KnowledgeAAAI Spring Symposium on Agent-Mediated

Knowledge Management (AMKM-03)

Sinuhé ArroyoInstitut für Informatik IFI

Next Generation Research Group University of Innsbruck, Austria

Juan Manuel DoderoComputer Science DepartmentUniversidad Carlos III de Madrid

Richard BenjaminsIntelligent Software Components

(iSOCO), S.A. Madrid, Spain

2

1. Introduction

2. Multi-agent collaborative production

Features and structure

Interaction within marts

Consolidation protocol

3. Case study

Course of the protocol

Results

4. Dynamics of markets

5. Conclusions

Dynamic Generation of Agent Communities from Distributed Production and Content-Driven Delivery of Knowledge

Dodero, Arroyo. — AMKM 2003 –3–

Multi-agentsystem

Casestudy

Conclusions

Intro

Dynamic of markets

1. Introduction

Collaborative knowledge management KM processes Distributed system Collaborative creation Task coordination needed

Creation or production Different interaction policies:

compete, cooperate, negotiate Structured interaction

Delivery Content-driven Communities of interest

delivery

production acquisition

Intro

Dodero, Arroyo. — AMKM 2003 –4–

Multi-agentsystem

Casestudy

Conclusions

Intro

Dynamic of markets

2. Multi-agent collaborative production

Producers’ collaboration (e.g. instructional designers) Asynchrony

• Development, exchange and evaluation of proposals are asynchronous.

• Different pace of creation Different levels of knowledge (Domain-level knowledge) Decision privileges (e.g. lecturers vs. assistants) Conflicts

Multi-agent architecture motivation Facilitates coordination when collaborating (e.g.,

compose a new educational resource) Allows different interaction styles (e.g., compete,

cooperate, or negotiate) Organizes interaction in distributed, but interconnected

domains of interaction

Multi-agentsystem

Dodero, Arroyo. — AMKM 2003 –5–

Multi-agentsystem

Casestudy

Conclusions

Intro

Dynamic of markets

System features

From a functional perspective… Consolidation of knowledge that

is produced From a structural

perspective… Multi-tiered structure Agents operate in tightly-

coupled hierarchical knowledge marts

Progressive consolidation of knowledge

From a behavioural perspective… Affiliation of agents into marts Evolution of marts

Multi-agentsystem

Interaction group S1 Interaction group S2

Interaction group M

Agent

Proxy Agent

Agent Agent

Agent Agent

Proxy Agent

Agent Agent

Dodero, Arroyo. — AMKM 2003 –6–

Multi-agentsystem

Casestudy

Conclusions

Intro

Dynamic of markets

Interaction within marts

Multi-agentsystem

Principles Agent rationality modeled as preference

relationships k1 > k2 or relevance functions u(k) Relevant aspects modeled as RDF triples (object,

attribute, value):• Submitter’s hierarchical level• Fulfilment of goals• Time-stamp

Message exchange Message types

• proposal ( knowledge, interaction )• consolidate ( knowledge, interaction )

Multicast, reliable transport facility

Dodero, Arroyo. — AMKM 2003 –7–

Multi-agentsystem

Casestudy

Conclusions

Intro

Dynamic of markets

Consolidation protocol

Distribution

Consolidation

Idle Failure

Success

any message

start(send proposal)

receive anyworse-evaluated

receive consolidation better-

evaluated

receive consolidation better-

evaluated

t0 expires receive proposal better-evaluated

receive anyworse-evaluated

t1 expires

Distribution

Consolidation

Multi-agentsystem

Dodero, Arroyo. — AMKM 2003 –8–

Multi-agentsystem

Casestudy

Conclusions

Intro

Dynamic of markets

3. Case study

Casestudy

Learning Object Course titled “Introduction to XML”

Roles 3 instructional designers, represented by agents

A1..A3 A1 is a docent coordinator

Task Development of the TOC A1 submits p, A2 submits q, A3 does nothing

Proposals p = Proposed manifest file with 6 chapters q = Modified manifest file, divides up chapter 5 in

two Evaluation criteria

Fulfillment of objectives Actor’s rank

Dodero, Arroyo. — AMKM 2003 –9–

Multi-agentsystem

Casestudy

Conclusions

Intro

Dynamic of markets

Course of the protocol

A1 A2

A3

Proposal p

Proposal p

A1 A2

A3

u(p) < u(q)Reply with q

Proposal q

Proposal qu(p) < u(q)

Start timeout t1

Start timeout t0

Start timeout t0

t0 expires

OK

A1 A2

A3

Consolidate q

Consolidate q

Start timeout t1

Termination:unsuccessful

OK

A1 A2

A3

Finish:successful

t1 expires

Initial exchange of proposals After receiving proposals

Consolidation after t0 expiration After t1 expiration

Proposal q

Casestudy

Dodero, Arroyo. — AMKM 2003 –10–

Multi-agentsystem

Casestudy

Conclusions

Intro

Dynamic of markets

Results: quality (grade of fulfilment)

75

5554

75

4343

75

5554

43

0

10

20

30

40

50

60

70

80

0 2 4 6 8 10 12 14 16 18 20

Proposals ordered by submission time

Gra

de

of

fulf

illm

en

t (%

)

Issued in two-mart scenario Issued in one-mart scenario

Casestudy

Dodero, Arroyo. — AMKM 2003 –11–

Multi-agentsystem

Casestudy

Conclusions

Intro

Dynamic of markets

Results: consolidation lifetime

1622 2183 3976 7401

190288

13439 7291

183027

0

20000

40000

60000

80000

100000

120000

140000

160000

180000

200000

1 2 3 4

Consolidated proposals ordered by instant of consolidation

Co

ns

olid

ati

on

life

tim

e (

tim

e u

nit

s)

Casestudy

Dodero, Arroyo. — AMKM 2003 –12–

Multi-agentsystem

Casestudy

Conclusions

Intro

Dynamic of markets

Results: number of conflicts

0,00

10,00

20,00

30,00

40,00

50,00

60,00

70,00

I1 A1 A3

Agents

No

. of

con

flic

ts/t

ime

un

it

No. of conflicts (two-mart) No. of conflicts (one-mart)

Casestudy

Dodero, Arroyo. — AMKM 2003 –13–

Multi-agentsystem

Casestudy

Conclusions

Intro

Dynamic of markets

4. Dynamics of markets

Dynamics of collaborative groups Agents affiliate to marts depending on the kind of

knowledge that they produce Marts evolve (merge or divide) depending on the

kind of knowledge consolidated within them Agents arrangement

Cognitive distance dk between agents and marts Defined from dissimilarity between issued

proposals’ attributes Agents operate in the nearest mart Agents relocate based on Knowledge production

Evolution of groups Mart fusion/division MajorClust algorithm

Dynamic of markets

Dodero, Arroyo. — AMKM 2003 –14–

Multi-agentsystem

Casestudy

Conclusions

Intro

Dynamic of markets

Dynamic of markets

Information brokering services Content-driven delivery Filters to deliver contents of interest Publish/subscribe pattern

Communities of users User agents subscribe to items of

interest User agents produce (publish) items Brokers’ routing tables are built Routing tables contain (hide) users’

layout into communities of interestDynamic of

markets

Dodero, Arroyo. — AMKM 2003 –15–

Multi-agentsystem

Casestudy

Conclusions

Intro

Dynamic of markets

Goal

Effective communications Reduce amount of info shared by brokers Reduce distance among agents and their

interested marts Evaluate

Mart’s optimal size Cost of agent’s relocation related to

brokers communication efforts Impact of mart’s evolution in the service

Find best clustering algorithm K-means, COBWEB, MajorClust,… etc

Dynamic of markets

Dodero, Arroyo. — AMKM 2003 –16–

Multi-agentsystem

Casestudy

Conclusions

Intro

Dynamic of markets

5. Conclusions

Features Bottom-up, multi-agent approach to collaborative

knowledge production systems Dynamic building of user communities Applicable to other collaborative KM production

tasks• e-Book & learning objects composition• Calendar organization• Software development (analysis & design)

Improvements Further validation in multi-tiered scenarios Test of mixed interaction styles (retract,

substitute, reject) Evaluation of dynamic evolution of martsConclusions