Communities with Participative. Knowledge Production and Content- driven Delivery. Practical Aspects of Knowledge Management (PAKM) 2002. Sinuhé Arroyo.
Dynamic Generation of User Communities with Participative Knowledge Production and Contentdriven Delivery Practical Aspects of Knowledge Management (PAKM) 2002
Sinuhé Arroyo Intelligent Software Components (iSOCO), S.A. Juan Manuel Dodero Departamento de Informática Universidad Carlos III de Madrid
1
Dynamic Generation of User Communities with
1. Introduction 2. Agent-mediated approach
Participative Knowledge
?
Structure of groups
Production and Content-
?
Interaction within groups
driven Delivery
?
An example
3. Dynamics of production groups ?
Distance-based group affiliation
?
Dynamic user communities
5. Evaluation: instructional design scenario
Arroyo & Dodero — PAKM 2002
?
Design of tests
?
Experimental results
4. Conclusions and future work –2–
Introduction ? Distributed knowledge management ? KM processes ? Distributed system
acquisition
production
? Collaborative, group level issues ? Coordination needed
? Production issues ? Different interaction policies:
compete, cooperate, negotiate ? Structured interaction ? Hierarchical, participative relationship
delivery participative
authoritarian
? Delivery issues ? Content-driven ? Communities of interest ? Dynamic spawning and evolution
Arroyo & Dodero — PAKM 2002
–3–
Agent-mediated approach ? Producers’ collaboration (e.g. instructional designers) ? Asynchrony • Non-frequent synchronous meetings • Management (development, exchange, evaluation) of products (e.g.
docs, exercises, software) is asynchronous
? Logically structured action • Domain-level knowledge (e.g. subject taught) • Decision privileges (e.g. lecturers vs. assistants) • Conflicts should be reduced
? Why multi-agent? ? To facilitate coordination of creation tasks (e.g., compose a new
educational resource) ? To help performing several interaction styles (e.g., compete, cooperate, negotiate, …) ? To structure interaction in separate, but interconnected groups
Arroyo & Dodero — PAKM 2002
–4–
System features ? From a static perspective… ? Function: consolidation of knowledge that is produced ? Static structure: multi-tiered • Agents operate in tightly-coupled interaction groups • Hierarchical structure of groups • Progressive consolidation of knowledge
? From an environment-relational perspective… ? Agent-mediated interaction ? Multi-agent architecture and coordination protocol
? From a dynamic perspective… ? User agents’ membership into groups ? Groups evolution: fusion and/or division
Arroyo & Dodero — PAKM 2002
–5–
Hierarchical structure: marts Interaction group M Agent Proxy
Agent
Interaction group S1 Agent
Agent Agent
Arroyo & Dodero — PAKM 2002
Interaction group S2
Agent
Agent
Proxy Agent
Agent
–6–
Interaction within marts ? Minimal requirements ? Agent rationality can be modelled as preference relationships
k2 or relevance functions u(k)
k1 >
• Non-despite of other, more powerful schemes (v.g., fuzzy logic)
? Relevant aspects can be modelled as pairs: • Possible attributes: submitter’s hierarchical level, fulfilment of goals,
time-stamps, etc. • Non-despite of other, more powerful representations (v.g., ontologies)
? Agent interaction by exchanging messages ? proposal ( object, interaction ) message (FIPA’s propose) ? consolidate ( object, interaction ) message (FIPA’s inform) ? Multicast & reliable transport facility
Arroyo & Dodero — PAKM 2002
–7–
Consolidation protocol ? Start of protocol Ai sends proposal(ki,n) to every agent ? Ai starts distribution phase and sets timeout t0 ?
? If Ai does not receive messages referred to n during t0... Ai sends a consolidate(ki,n) to every agent ? Ai starts the consolidation phase ?
? If Ai receives a proposal(kj,n) from A j / kiki, before t1 expires, then the protocol finishes unsuccessfully.
?
? When Ai’s timeout t1 expires, the protocol finishes successfully
Arroyo & Dodero — PAKM 2002
–8–
Consolidation protocol receive any worse-evaluated start (send proposal)
Distribution Distribution
receive consolidation better-evaluated
any message
Idle
receive any worse-evaluated
t0 expires
receive proposal better-evaluated
Consolidation Consolidation t1 expires
Arroyo & Dodero — PAKM 2002
Failure receive consolidation better-evaluated
Success
–9–
Instructional design example
proposed modification: divide up chapter 5
Arroyo & Dodero — PAKM 2002
–10–
Example: course of the interaction Proposal p
Start timeout t0 A1
Proposal q
Proposal p
Start timeout t0
Proposal q
A2
A1
Proposal q
A2
u(p) < u(q) Start timeout t1
A3
u(p) < u(q) Reply with q
A3
OK Initial exchange of proposals
t0 expires
After receiving proposals
Termination: unsuccessful A1
Consolidate q A3
A2
Start timeout t1
Consolidate q
OK Consolidation after t0 expiration Arroyo & Dodero — PAKM 2002
Finish: successful A1
A2
t1 expires A3 After t1 expiration –11–
Dynamics of production groups ? Requirements ? 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.
? Distance-based clustering ? Cognitive distance dk between agents and marts • dk(agent1, agent2) • dk(agent, mart) • dk(mart1, mart2)
? Built from dissimilarity between proposals’ attributes issued by
agents
? Agents are arranged to operate in the nearest mart ? Mart fusion/division ? K-clustering algorithm (v.g., COBWEB, K-SODATA) Arroyo & Dodero — PAKM 2002
–12–
Dynamic user communities ? 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 interest
? How to build dynamically… ? Monitor cognitive distances ? Apply k-clustering algorithm ? Re-organize membership of agents
Arroyo & Dodero — PAKM 2002
–13–
Evaluation scenario ? Learning Objects ? Curriculum units that can be
assembled to build higher-level learning contents. ? IMS/SCORM specifications ? XML manifest file ? Coordination needed
? Issues in collaboration ? Conditioned by partners’
competency ? Speed vs. quality tradeoff ? Avoid duplication of effort
Arroyo & Dodero — PAKM 2002
–14–
Design of tests ? Task ? Development of a Learning Object (TOC of a course on XML)
? Roles ? Docent coordinators + instructional designers + students
? Test scenarios ? one-mart (docent coordinators + instructional designers) ? two-mart (instructional designers + students)
? Proposals’ evaluation criteria ? Fulfilment of educational objectives + Rank + Time-stamp
? Measures ? Quality of the L.O. ? Stability of consolidated objects ? Conflicts decrease
Arroyo & Dodero — PAKM 2002
–15–
Quality (grade of fulfilment) Issued in two-mart scenario
Issued in one-mart scenario
80 75
75
75
Grade of fulfillment (%)
70 60
55
54
54
55
50 43
43
40
43
30 20 10 0 0
2
4
6
8
10
12
14
16
18
20
Proposals ordered by submission time Arroyo & Dodero — PAKM 2002
–16–
Consolidation lifetime
190288
Consolidation lifetime (time units)
200000
183027
180000 160000 140000 120000 100000 80000 60000 40000 20000
13439
7291
0
1622
2183
1
2
7401
3976 3
4
Consolidated proposals ordered by instant of consolidation
Arroyo & Dodero — PAKM 2002
–17–
Number of conflicts No. of conflicts (two-mart)
No. of conflicts (one-mart)
No. of conflicts/time unit
70,00 60,00 50,00 40,00 30,00 20,00 10,00 0,00 I1
A1
A3
Agents Arroyo & Dodero — PAKM 2002
–18–
Conclusions and future work ? Features ? Bottom-up, multi-agent approach to collaborative knowledge
production systems ? Dynamic building of user communities ? Applicable to several CSCW tasks • e-Book composition • Calendar organization
• Software development (analysis & design)
? Improvements needed ? Further validation in participative scenarios ? Mixed interaction styles to be tested ? Quantitative evaluation for dynamic evolution of marts
Arroyo & Dodero — PAKM 2002
–19–