From: AAAI Technical Report SS-02-02. Compilation copyright © 2002, AAAI (www.aaai.org). All rights reserved.
A Multiagent Frameworkfor Collaborative Conceptual Learning
Using a Dempster-Shafer Belief System
Leen-Kiat Soh
Computer
Scienceand Engineering
Universityof Nebraska
115Ferguson
Hall
Lincoln, NE
(402) 472-6738
lksoh@cse.unl.edu
Abstract
In this paper, wedescribea multiagentframework
for collaborativeconceptuallearningusing a Dempster-Shafer
belief systemin the domainof informationretrieval. In our
multiagentsystem,each agentmaintainsa databaseof documents,entertainsdifferent queriesfromits users, andthus
learns a uniquedictionaryof concepts.Filed for eachconcept is a set of keywords
collectedfromthe documents
supportingthat concept.Adocument
maybe filed undervarious
conceptsand thus conceptsmayshare keywords.This providesfor a metricto evaluatethe relevanceof a document
to
a query.Ourproposedworkenablesthe queryof an agentto
be composed
at a conceptuallevel or expandedat a conceptual level, justifiedbya set of keywords,
andto be learnedby
other agents throughcommunications.
Thelearnedconcepts
are stored in eachagent’suniquetranslation table. In this
manner,agents are able to evolve independentlytheir own
knowledge
whilemaintainingtranslation tables throughcollaborativelearningto helpsustainthe information
retrieval
process.
Introduction
The goals of our research are to (1) promoteunderstanding
amongagents of a community,thus reducing communication costs and inter-agent traffic, (2) improvecooperation
amongneighbors of a community, thus enhancing the
strength (productivity, effectiveness, efficiency) of
neighborhoodand supporting the distributed effort of the
community,(3) encourage pluralism and decentralization
within a multi-agent community--specializationof agents
of a community
since each agent can rely on its neighbors
for tasks not covered by its capabilities, and (4) enable
collaborative learning to improvethe throughput of the
community,the intelligence in communicationand task
allocation, the self-organization within the community,
and
integrity of the community.To address these goals, we
define a multiagentframeworkfor collaborative conceptual
learning using a Dempster-Shaferbelief system, with an
example domain in information retrieval (Frakes and
Baeza-Yates1992).
In our system, each agent maintains a database of documents and handles different queries from its users. As a
result, each agent learns its ownconceptsbased on its experiences, functional utilities, and purposes. Duringthe
training phase, the user submits a documentwith a set of
labels (or conceptual terms). The agent receives the submission, processes the document,and stores it in its database. The labels or conceptsare then matchedto the existing concept database, and the keywordsextracted from the
documentsare matchedto the keyworddatabase. By analyzing the two databases, the agent is able to absorb the
concept(newor modified)and what it meansinto its knowledge base. During the execution phase, each agent is an
information retrieval system, handling queries from users.
Whenan agent does not recognize a query, it relays the
query to one of its neighbors. It is during these relays or
exchangethat different conceptsare shared. In both phases,
knowledgeis incrementallylearned and this has to be consistently supported. Towardsthis end, we use a DempsterShafer belief system(Shafer 1976).
In this paper, we first present the methodologyof our
framework:conceptlearning, queryprocessing, translation,
interpretation, action planning, and query composition.
Then we will concentrate on the application of the belief
systemto facilitate learning--incrementalwithin an agent,
and collaborative amongagents. Thenwe will discuss plausible applications for the proposedframeworkbefore concluding.
Ourdiscussion here is related to (Williamsand Tsatsoulis
1999). In (Williams and Tsatsoulis 1999), however,
agents werenot able to learn collaboratively in a multiagent
system. Instead, the learning was conductedonly between
two agents via exchangeof concepts (ontologies) wherethe
agents were neither able to adapt to changes in concept
definitions nor able to handle multiple assertions fromdifferent neighbors. Moreover, our framework addresses
translation and interpretation of concepts, queryprocessing
and compositionfor informationretrieval and task distribution amongagents, and action planningbased on traffic and
agent activities, whichindirectly control the learning rates
of the agents.
Copyright
0 2002,
American
Association
forArtificial
Intelligence
(www.aaai.org).
Allrights
reserved.
Methodology
In our framework,each agent is able to learn concepts and
distribute queries, in addition to the normalinformation
retrieval tasks. Thefirst capability cultivates the intelligence of the agent while the second enables cooperation
and communicationamongagents in the community.Together, these twocapabilities facilitates incremental,collaborative learning.
Concept Learning
Ouragent uses a list representationto organizeits concepts
into a hierarchyand a set of rules to describeeach concept.
For example,if the user defines under Sports
three categories: Football,Basketball,
and Tennis,and
further defines under Football
two categories: NFLand
College, then the list representation of the conceptual
hierarchyis
(Sports(FootballNFL College)Basketball
Tennis
)
Fromthe list, an agent identifies the leaf nodes, i.e.,
NFL,College,Basketball,
and Tennis.These are
what we call basic concepts. Eachconcept is supported by
a set of keywordsextracted from processing the documents
that exemplifythe concept.
Theconcept learning process is self-motivated and will
be invokedwhenthere is a noticeable changeto the database (e.g., a user submission).Eachagent computesa discrepancy percentage that is based on the differences in
tokens of the old and newrepresentation lists and the percentage of increase in the numberof examplesfor a partitular concept. Whenthis discrepancypercentageexceeds
a certain threshold, the agent automatically invokes the
concept learning phase, which involves tokenization and
rule building. Thetokenization algorithm is a combination
of conditional frequency of occurrences and rank-based
filtering. At the end of the tokenization process, each
documentwill be represented with a set of keywords.
Since each documentfalls under a set of concepts, the associated set of keywordsbecomesan example that describes each of those concepts. For each document,there
is a set of keywords. Combiningall documentsfor one
concept, we can establish frequency ranges for each keyword. Giventhe set of all keywordsdescribing a concept,
webuild a vector consisting of the identifier of the document(object name), the frequencyof occurrencesfor each
keyword(attribute values), and the concept it belongs
(the class). Toconstructrules, wefeed the vector field into
an inductive learner. Thelearner parses the vectors into a
decision tree that deterministically allocates each example
into a semantically unique branch. Brancheswill then be
traversed--attribute values extracted and comparatives
introduced--toarrive at rules. Anexamplerule is:
If(university > 0.2000) and
(nebraska > 0.1500) and
(lincoln > 0.0320) an4
(washington < 0.0003) then
NU.
10
The above rule says that if the keyword "university"
appears in the documentmore than 20%, and the keyword
"nebraska" appears in the documentmore than 15%, and
the keyword "lincoln" appears in the document more
than 3.2 percent, and keyword"washington" appears in
the documentless than 0.03%, then the concept is ’2~."
This rule thus is used in query processing: if a query asks
for documents under the concept "NU," the agent knows
howto evaluate the relevance of the documentsin its databaseto the query.
Query Processing
Queryprocessing allows the knowledgeand distributed intelligence of an agent be maintainedlocally. It also allows
the user to submit a concept with its exampledocuments.
Figure 1 showsthe modulesthat an agent uses to process an
incomingquery. Note that a query can be submitted by a
user or another agent. In this discussion, we will focus on
the latter.
FigureI Thequeryprocessingfunctionalityof an agent
A queryis a compositionof the querying agent’s identifier, the queried agent’s identifier, the conceptname,the
rules that define the concept, the description of the request
(relevance degree, the numberof documents,etc.), and the
query’s originator. Whena query from another agent comes
in, the agent first attempts to translate the embedded
concept. If a translation is found,then the agent retrieves relevant documentsand sends it to the querying agent. Otherwise, the conceptis passed to the interpretation module.If
the conceptrequested by the queryis similar or relevant to
what the agent knows, then it duly supplies the querying
agent with the relevant documents.If both the interpretation
and translation fail to recognizethe concept, then the agent
plans its next actions basedon the traffic and activity situations at its neighborhood(Figure 2). If the neighborhood
traffic is not congestedand the agent itself is not busywith
its owntasks, then it will adopt the queryand ask help from
other agents. Otherwise,it will reply a NULL
to the querying agent, effectivelyterminatingthe query;or if the agent is
busybut the queryis of highpriority, it will voluntarilysupply the information regarding where the knowledgemaybe
solicited, practically re-directing the responsibility backto
the queryingagent.
Oncethe queried agent has decidedto help, it will compose a query or queries based on what its neighborhood
monitor has observed and learned so far. Issues such as
specialties, personalities, and communication
experiencesof
the neighboring agents determine howa query should be
composed.The agent might want to send a query to a par-
ticular neighbor with high priority, or broadcast the query
to every neighbor with low priority, or target a group of
specialized neighbors, or ask for help from a helpful and
knowledgeable agent. This decision making process is
based on the observations that the agent has experienced
and the lessons it has learned interacting with its neighbors.
This approach enhances the cooperative efficiency and
effectiveness between an agent and its neighbors, thus
strengthening the collaboration within the system.
Translation
Whenthe translator receives a query, it extracts the concepts and parses the example documents to obtain rules.
Suppose that there is only one concept, "NU," and the rules
(after parsing the documents)are:
If (university > 0.2000) and
(nebraska > 0.1500) and
(lincoln > 0.0320) and
(washington < 0. 0003) then
NU.
The semantic collected the rule is:
(univer s i ty > 0.2000 nebraska
0.1500 lincoln > 0.0320 washington
O. 0003 ),
<
where a semantic component is, for example, ’~niversity > 0. 2000", a semantic token is "university",
a semantic conditional frequency of occurrences (or semantic frequency) is "0. 2000", and, finally, a semantic
comparative is ">".
Each agent maintains a comprehensive translation table.
Each table lists the concepts that the agent knowsand maps
them to the corresponding concepts of the neighboring
agents. Table 1 shows an example of translation tables.
Table 1 says, for example, the following: "Hy concept
NU is relevant to the concept Campus of
Neighbor1 with a credibility of 2.39",
’~4y concept Monet is relevant to the
concept Oil Painting of Neighbor5 with
a credibilityof 2.97", etc. The calculationof
the credibility value will be discussed in the interpretation
section. Only translations that are credible will be recorded in the table. Wewill discuss credibility later in this
paper.
Table 1 Translation table
CONCEPT
NU
Monet
Sports
AI
N1
Campus
2.39
N2
N3
NU4.10
N5
Nebraska
3.12
Art 2.76 European Oil PaintArtists
ing 2.97
2.28
ESPN3.61 NCAA
NCAA
Sports
1.99
2.56
News3.41
Artit~ial
Research Thesis
AI3.55
Intelligence 2.57
2.12
3.91
11
In the beginning, each agent has an empty repository of
translation tables. At the birth of an agent, an agent learns
from the users’ submission (or queries). Then the agent
learns about the relations it has with its neighbors through
two functional occasions. First, when it queries another
agent for certain knowledge or information. Second, when
it receives a query from another agent. Whenan agent queries another agent for certain knowledgeand if the queried
agent responds positively with its ownsemantics, the querying agent will duly interpret it and update it in its translation
table. Whenan agent receives a query from another agent,
if it does not have a readily available and up-to-date translation, then the agent interprets the semantics that accompany
the query. At the end of the interpretation, if the agent is
able to recognize the semantics, it then reflects the learned
mappingsin the translation tables.
The interpretation-driven learning will be discussed in
further details later.
Refinements. The concepts of an agent are constantly being modified--narrowed, expanded, etc. To accommodate
this evolving behavior, the translation tables must facilitate
refinements.
Whena re-learning occurs, a new hierarchy and sets of
new rules ensue. The agent automatically examines the
translation tables to update the changes. Suppose the agent
has a concept called NU. If this concept’s old keywords are
the same as the newcounterparts, then the translations related to NUare preserved. If they differ, then the agent refines the credibility value of the translation based on the
degree of difference between the two sets of keywords
based on a mass measure. The agent then computes a refinement factor, the average for the mass measurements of
all possible pairings between the new set of keywords and
the old. This refinement factor is used to adjust the credibility value of each translation. If any of those translations’
credibility drops below an acceptable threshold, then that
translation is removed from the table. This new evidential
refinement approach allows (I) the preservation of previously learned translations,
and (2) the incorporation
newlylearned semantics into the existing translations.
Interpretation
Whena query falls to be matched through during translation, it is forwarded to the interpretation step. Whenmatching the semantics of the query with the semantics of a resident rule, if the two sets are synonymic, then a mass of 1.0
will be added to the interpretation
scores. Whenwe say
’synonymic’, we refer to the inclusiveness and exclusiveness
when comparatives are involved. For example, if the querying semantic has a semantic component ’~niversity
>
0. 2000" and the resident semantic has "university >
0.1500", then the resident semantic component is said to
be inclusive of the querying semantic component. A similar
observation can be said about the less-than comparative.
Hence, two semantics are synonymic when the resident
components include the querying components. In the sense
of rule-based systems, if a semantic is matched, then the
resident rule is fired by asserting the conceptnameentailed
by the resident rule with a mass of 1.0. The translation
betweenthe concepts that describe the synonymicsemantics will thenbe recordedin the translation tables.
Ouragent is also equippedto handle relevant matching
since a synonymicmatching is rare. Suppose we have a
querying semantic "(college > 0. 1300 nebraska
> 0.1237cornhusker> 0.1000)"and the resident semantic"(cornhusker
> 0. 2000 university > 0.1200 nebraska > 0.1138)". Now,
thefirstsemantic
component
is notmatched.
Thesecond
semantic componentis matched. The third semantic component is partially matched:the semantictoken is matched
but the comparativerequirementis not fulfilled. Suppose
the numberof resident semantic componentsis Nsc, the
massof the assertion of the abovepartially-matchedrule is
~ matched(residentcon compi , query con)
Mass=iffil
N$¢
If a resident semanticcomponent
is matched,then the function matchedreturns 1.0. If the semantic token of a resident semantic componentis not found in the querying semantic, then the function matchedreturns 0.0. If the semantic token matches but its semantic frequency is excluded (fromthe range), then matchedreturns
1.0-J resident con freq-query con freq {
Thus, in the exampleabove, the mass of the match is
0.6634. Andwe say the concept names of the two semantics are probably relevant with a massof 0.6634. Whena
rule is matched,either completelyor partially, that rule
asserts the conceptnameto the interpretation knowledge.
Action Planning
Our agent has an action planner, as shownin Figure 2.
a#
Figure2 Theactionplanningandquerycomposition
of anagent
Wewill briefly discussits activities here. Anaction
plannermonitorsits environment
to decidewhichcourseof
actionsto take: (I) offer help by adoptingandrelaying
12
query to other agents, (2) supply the necessarytranslation
informationand let the queryingagent retrieves the knowledge itself, or (3) return a NULL
responseto terminate the
query. Such a decision is based on two criteria observed
from the environmentand the priority of the query: (1) the
activities the agent currentlyis engagedin, and (2) the traffic situation amongthe agent and its neighbors.Bothpieces
of informationare providedby monitors--the activity monitor and the traffic monitor.Thequerypriority signifies the
importanceof the request to the queryingagent.
Query Composition
As shownin Figure 2, when an agent decides to relay a
query, it needs to composeits owninterpretation of the
query. The query composerrefers to the neighborhoodprofiler to constructits queries. Thisprofiler compilesa list of
experience-based
characterizationsof specialties, personalities, and communication
experiences of the neighbors. For
every interaction betweenan agent and its neighbor, the
successof the interaction, the accuracyof the interpretation
and translation, the usefulness of the knowledgeand information, the responsivenessof the neighbor,and other interagent issues are logged and evaluated. Eachcriterion will
be scored; the scores averaged;and all idly-linked neighbors
ranked in terms of the average. The composerthus constructs its queries based on the average: low priority to
friendly neighbors, high priority to unresponsiveneighbors,
etc. The composerthus determines general and selective
broadcasts and one-to-one communication.
QueryAcknowledgement.
After a query is processed, the
queriedagent returns the result backto the queryingagent.
The querying agent has to acknowledgeit. Figure 3 shows
the modulesinvolved in query acknowledgement.
Whenthe result is received, the incident will be recorded
in the knowledgebase regarding the neighborhood:(a) success of the interaction---based on the NULL,
information,
and knowledgetyping if the response, (b) responsiveness
the neighbor--basedon the elapsed time betweenquery sent
and response received, (c) nature of the query--concept
categories, priority levels, originators, designators,etc., and
(d) log of activities and traffic observations.Theknowledge
base provides a library from which an agent can learn to
composebetter queries as the community
evolves.
The knowledgemodifier maintains knowledgebases of
concepts, translation tables, and neighborhoodbehavioral
observations. It also compilesstatistics such as average,
risk factor, friendliness factor, and successrate and records
themonto the appropriate knowledgebases.
I- Z m] (~)n2(~)’
lxlI~s,
where r ~e 0, and[mI ~ m2 ]{O)= 0.
To showhowwe apply Dempster-Shaferbelief system to
our interpretation process, let us consider a simple case.
Supposeafter matchingthe semantics to our rule base, we
arrive at two assertions: ~ with mass 0.7 and Monetwith
mass0.2. Hence, ~ correspondsto our belief:
INU} 0.7
O 0.3
and m2 correspondsto our belief:
,~w#w
{ Monet] 0.2
O 0.8
Figure3 Thequeryacknowledgement
functionalityof an agent
Then we can compute their combination m3 using the
rule of combinationabove(presented as a table below):
|Mo~) 0.2
[NU)
O.7
IMonet, NU}0.14
e 0.8
[NU]
Dempster-Shafer Belief System
Anagent performstwo types of learning. It learns incrementally, refining its concepts wheneverthere is a new
submission. It also learns collaboratively, refining its
translation table wheneverthere is a querythat promptsthe
agent to ask for help from its neighbors. The underlying
problemis howto combinethe various assertions madeby
the relevant matchingthat wediscussedearlier in a consistent manner. Towardsthis end, we incorporate the Dempster-Shafer theory (Shafer 1976)for building a belief system that receives evidenceand maintainsglobal beliefs in
its assertionsconsistently.
Supposethat all the concept namesthat an agent understands, stored in its ontologies, are of the frameof discernn~nt or universe U. A proposition in favor of a concept name,F, is thus an assertion as previously described.
Thus, the set of all propositions is P(U), the powerset
U. Let m: P(U)---) [0,1] be a function--a basic probability
assignment--satisfying conditions for a certainly false
proposition, m(O)=0, and for a certainly true vroposition,
~m(F)=l. The belief function, Bel:P(U)--)[O,1], is
d’eghed in terms of the basic probability assignment m:
Bel(r)= ~ m((~). This tells us the degree of belief associated wit#g~eproposition r as the probability massassociated with r and its subsets. Theplausibility of a proposition is further defined as Pls(F)=l-Bel(-~I’). Hencea
proposition is alwaysboundby [8el, Pls] in terms of the
confidencein its perceivedtruthfulness. To combinevarious pieces of evidence for building up beliefs in favor of
various propositions, the Dempster’srule of combinationis
used. Supposewe are given two assignments (two pieces
of evidence), I and m2, a nd we want t o c ombine them
into a single piece of evidence. Hence,we compute
13
0.$
[Monet) 0.10
0.56
0.40
In this manner,all newevidencewill be incorporatedinto
the previously accumulatedbeliefs consistently. Fromm,,
wecan further computethe evidential interval, [Bel, Pls],
for eachof the conceptnames.For I~, the interval is [0.56,
0.90]; for Monet,the interval is [0.I0, 0.44].
ConceptDisambiguation
Until now,after the rule-hased, semantics-drivenassertions
and the evidential combination,wearrive at a set of evidential intervals for the propositions or concept names.For
example, supposewe have the following:
CONCEPT
I
Sports
AI
Monet
NU
EVIDENTIAL INTERVAL
I
[0.007968,
0.079682]
[0.107570, 0.179283]
[0.286852, 0.358566]
I0.525896, 0.597610]
In the above example, the concept NUhas the highest
credibility (discussed later) to be true, followedby Monet,
AI,and Sports.
Hence, we have a fuzzy or ambiguous
understandingof the querying concept. For our disambiguation process, we followthe twoaxiomsof evidential interval
analysis:
Axiom1 The higher the belief and the plausibility values,
the morecredible the propositionis.
Axiom2 The closer the belief value is to the plausibility
value, the morecredible the propositionis.
AxiomI follows naturally from the work of the Dempster-Shafer theory. Onthe other hand, Axiom2 punishes
ignorance.That is, if the agent thinks that a propositionis
very plausible but believes with little confidencethat the
proposition is true, then the agent is ignorant about the
proposition. Following from the above two axioms, we
devise a measureof credibility of a propositionas:
Credibility(I" )= PIs(F)+Bel(F
PIs(F)-BeI(F)"
Duringinterpretation, the conceptthat yields the highest
credibility will be the winningconcept. Notethat if the
credibility of the winning concept is below a certain
threshold, then the interpreter realizes that it doesnot understand or recognizethe queryingconcept. This provision
preventslow-qualityrecognition.
At the end of this stage, the interpreter performsone of
the following: (1) If the winningconceptpasses the credibility test, then the agent supplies the queryingagent with
what it knows,i.e., the documentsunder the winningconcept. Thetranslation will also be recordedin the translation tables, or (2) If the winningconceptfails the credibility test, then the agent turns to the translator moduleof its
system.
Concept Amalgamation
This process is triggered by the combinationof (I) the lack
of a credible winningconcept, and (2) the existence of
credible, relevant non-singleton concept structure. The
objective is to promotenon-singleton sets of concepts,
such as {Basketball, NU},to a recognizableconceptstructure. For example, suppose the set {Basketball, NU}has
an evidential interval of [0.6 0.9]. Its credibility is 5.0.
Supposethis passes our filter and the winning concept
fails.
The amalgamationprocess will register the conceptual
complex[Basketball, NU}to a complex-relevant table,
together with the semanticsthat support the complex.Further, it will recordthe translation andits credibility to the
correspondingspace under the querying agent.
This provision self-motivates every agent to build and
learn complexconcepts, whichin turns increases the complexity and level of understanding amongthe agents with
distributed ontologies.
Belief System and Incremental Learning
Ouragent conductsincrementallearning at different times.
Duringthe training phase, the agent handles user submissions, performsinductive learning to obtain semanticrules,
and builds its concept database. Since these submissions
or documentsare received sequentially, the learning is incremental. During the execution phase, users can submit
both queries and new documentsto an agent. Whennew
documents
are submitted,an agent revises its conceptdatabase through the belief systemin the following manner.If
14
the newdocumentsare submitted with an existing concept
name,then the agent essentially finds a translation between
the concept name with new documents(evidence) and the
concept namewith old documents. The credibility of the
translation is then used to re-weight massof the existing
semantic rules and the new rules derived from the new
documents.This revision is then propagatedto all related
translations in the table. Onthe other hand, whena queryis
submitted and the agent fails to recognize the concept of
that query, it mayrelay it to other agents. If anotheragent
retrieves the relevant documentsand returns them, then the
originating agent learns that newconceptusing the retrieved
documentsas examples, absorbing it into its concept base
via belief systemin a mannersimilar to the aforementioned.
Belief System and Collaborative Learning
Our multiagent frameworkis a distributed information retrieval system. The agents learn to re-direct tasks and reformulate queries for better retrieval. The fundamental
mechanism
that enables such behavior is through the maintenanceof the translation tables in the system.Thepresence
of a unique translation table at each agent increases the
autonomyof the agents in the system, allowingeach to specialize for specific sets of queries and documents.Further,
an agent is designedto relay a querythat it cannotsatisfy to
other agents for help. Anagent maynot satisfy a query
whenit does not recognize the query concept, or does not
find a relevant match, or does not retrieve documentswith
high relevance values, or does not find enoughdocuments
as required. As a result, the agent can use its translation
table to locate useful neighborsto approachfor help. If the
conceptquerydoes not havea translation but the agent does
have a few lowiy-relevant documents,then it will supply
those to its neighborsas examples.The key of our utilization of collaborative learning is that each agent has its own
set of conceptsto facilitate retrieval accuracyand speed.
Communications
are only necessary whenagents need help.
Throughcommunications,queries are relayed and concepts
are shared. Thus, the agents only learn necessary translations basedon their experiences, makingthe learning process efficient and effective. The collaborative learning results are essentiallythe translationtables.
Applications
Our frameworkacknowledgesthe essential ontological diversity that alwaysexists amongagents of a diverse communitydue to different utilities (Geneserethand Nilsson
1987). It encouragesthe growthof such as a community-not by restricting whatthe dictionary shouldbe---instead by
promoting the uniqueness and freedom of expression of
each memberthrough co-operative learning in a multiagent
framework.Second, previous research has focused on using
a pre-defined, commonontology to share knowledgebetweenagents by using a common
set of ontology description
primitives such as KIF (Genesereth and Fikes 1992) and
Ontolingua(Gruber 1993). However,the approach of using
global ontoiogies has problems due to the multiple and
diverse needs of agents and the evolving nature of ontologies (Mineau 1992). Third, agents mayhave disparate
references, whichlead themto refer to the sameobject or
conceptusing different terms and viewpoints,i.e., diverse
ontologies (Bond and Gasser 1988). Our multiagent
frameworkallows membersor agents to learn and identify
what these disparate references mean--constituting the
distributed intelligence and collaborative learning of the
agents.
The followingidentifies the potential applications that
wewill investigate:
1. Conceptual Retrieval--Instead of pure string matching, we can nowperformconceptualretrieval based on
the ontologies learned from interacting with other
agents. This will help in the creation of a personalized, conceptualminer.
2. Agent Personalization--An agent can be a personal
assistant to its user in locating relevant documents
that
are conceptually synonymicor relevant to what the
user specifies. Anagent can help tailor the content delivered to the user by interpreting and translating it
automatically, extracting importantcorporal terms for
inspection.
3. Assistant to Web-searchengines---Our agent can provide ontologies to improvethe precision and recall
speed of searching on the Web.It can find related
words or terms and supplement search engines for
query expansions. Our muitiagent framework can
providean infrastructure for a distributed, specialized
Web-searchengine. That is, each agent in the system
is specialized in retrieving a certain domainof information.
Infrastructure---With the mul4. Distributed Knowledge
tiagent design strategy outlined in this proposal, our
multiagentsystemcan act as the infrastructure to distributed knowledgeand knowledgesharing. This undoubtedly provides a structured corridor to mining,
consuming, and disseminating the Web, which is a
semi-structured, decentralized environment.In addition, our proposedsystemprovides a practical testbed
for dealing with agent interaction issues between
communitiesand within communities. The inter- and
intra-communityinteractions are vital to understand
howdistributed knowledgeinfrastructure should be
constructedto obtain effective and efficient knowledge
sharing.
5. Ontology Managerand OrganizermHierarchiesof the
concepts together with synonymand relevance networksof the translations can be organized as conceptual graphs (Sowa1984) to represent corporate understanding about domainconcepts and typologies. The
networks can be used within a communityto help
managinginformation flow and technology transfer.
This allows the evolution of a global ontologyvocabulary (Gruber and Oisen 1994) of a communityand
possible a polarization of such vocabularywithin the
15
.
amorphousuniverse of communities.A possible result
is a pseudo-thesaurusthat documentsthe translations of
different languages. For example, one might use English, German,or Chinese(HanYuPin Yin) to namethe
semantic concepts group the documents.Assumingthat
these users are responsible and careful, theoretically,
our agent is able to obtain a conceptualtranslation between an English term (Oil Painting) and a Chinese
term (You Hua), for example, if the two users bookmark the similar Webpages.Finally, conceptual taxonomies,or ontologies, can be useful for indexing and
organizinginformation and for managingthe resolution
of conflicting defaults (Woods1991).
Intelligent Multiagent Systenv--Our agent provides
modelsfor researchersin multiagentsystemsto investigate how, within a community,neighborhoods form:
basedon friendliness of the interacting neighbors,specialties of the neighbors,helpfulness of the neighbors,
communication
costs, etc. It is a useful tool to observe
various population-relatedbehaviorssuch as isolation,
defragmentation,co-operation, specialization, domination, alignmentand re-alignment, collaboration habits,
etc.
Conclusions
Wehave described a multiagent frameworkfor collaborative conceptual learning using a Dempster-Shaferbelief
system for information retrieval. This frameworkconsists
of conceptlearning, queryprocessing, translation, interpretation, action planning, and query composition. The concept learning is inductive and the learned knowledge
is represented in semanticrules. Thequeryprocessingfirst tries
to translate incomingquery, and if that fails, attempts to
interpret it. A translation is successful whena synonymic
matchis found. Aninterpretation is successful whena relevant matchis found. To help with the matching,each agent
maintainsa translation table, identifyingits ontologicalrelationships with its neighbors.Eachagent has an action planner to decide whetherto adopta query, to reject a query, or
to process a query. Thedecision is based on the agent and
traffic activities that the agent monitors.Finally, whenan
agent decidesto ask for help fromits neighbors,it composes
its querybasedon its profiles of the neighbors.Thisutilitybased evaluation allows the query composerto determine
general and selective broadcasts and one-to-one communication. Wethen apply the Dempster-Shaferbelief systemto
our agents" interpretation process. The belief systemallows
the agent (1) to learn newexampledocumentsand adjust its
semantic rules incrementally, (2) to learn the translation
table collaboratively, and (3) to computethe credibility of
translation (or a proposition). Thus, this belief systemprovides a mechanismfor collaborative learning amongthe
agents. Our proposedworkenables the query of an agent to
be composedat a conceptual level or expandedat a conceptual level, justified by a set of keywords,and to be learned
by other agents through communications.The learned concepts are stored in each agent’s uniquetranslation table. In
this manner,agents are able to evolve independentlytheir
own knowledge while maintaining translation tables
throughcollaborative learning to help sustain the information retrieval process. Withthe collaborative learning using the Dempster-Shaferbelief system, our multiagent
frameworkaffords high autonomy,flexibility, and specificity for each agent--the agents can learn fromeach other to
identify semanticallysimilar concepts(even if the concept
namesare completely irrelevant) and can relay queries
amongagents based on the translation for distributed informationretrieval.
References
Bond, A. H. and Gasser, L. (eds.) 1988. Readingsin Distributed Artificial Intelligence, San Mateo, CA:Morgan
Kanfmann.
Frakes, W.B. and Baeza-Yates, R. (eds.) 1992. Information Retrieval: Data Structures and Algorithms, Upper
SaddleRiver, NJ: Prentice Hall.
C-enesereth, M. and Nilsson, N. 1987. Logical Foundations of Artificial Intelligence, Palo Alto, CA:Morgan
Kanfmann.
16
Genesereth, M. and Fikes, R. 1992. KnowledgeInterchange
FormatManualVersion 3.0, Technical Report Logic-92-1,
Stanford LogicGroup,Stanford University.
Gruber, T. R. and Olsen, G. R. 1994. An Ontologyfor Engineering Mathematics, Proceedings of the 4th International Conferenceon Principles of KnowledgeRepresentation and Reasoning (KR94), Bonn, Germany,May24-27.
MorganKanfmann,258-269.
Mineau, G. 1992. Sharing Knowledge:Starting with the
Integration of Vocabularies, in H. Pfeiffer and T. Nagle
(eds.), ConceptualStructure: Theory and Implementation,
Proceedingsof the Seventh Annual Workshop,Las Cruces,
NM,July 8-10, Springer-Verlag, 34-45.
Sharer, G. 1976. A Mathematical Theory of Evidence,
Princeton, HJ: Princeton UniversityPress.
Sowa,J. 1984. ConceptualStructures: Information Processing in Mind and Machine, Reading, MA:AddisonWesley.
Williams, A. B. and Tsatsoulis, C. 1999. Diverse WebOntologies: What Intelligent Agents Must Teach to Each
Other, WorkingNotes of the AAAISpring SymposiumSeries
on Intelligent Agents in Cyberspace,Stanford, CA,Mar2224, 115-120.