CONCEPTS IN ACTION: PERFORMANCE STUDY OF AGENTS

LEARNING ONTOLOGY CONCEPTS FROM PEER AGENTS

Leila Safari

, Mohsen Afsharchi

and Behrouz H. Far

Department of Electrical and Computer Engineering,University of Zanjan, Zanjan, Iran

Department of Electrical and Computer Engineering, University of Calgary, Calgary, Canada

Keywords:

Ontology, Concept learning, Multi-agent communication.

Abstract:

The ability to share knowledge is a necessity for agents in order to achieve both group and individual goals. To

grant this ability many researchers have assumed to not only establish a common language among agents but

a complete common understanding of all the concepts the agents communicate about. But these assumptions

are often too strong or unrealistic. In this paper we present a comprehensive study of performance of agents

learning ontology concepts from peer agents. Our methodology allows agents that are not sharing a common

ontology to establish common grounds on concepts known only to some of them, when these common grounds

are needed by learning the concepts. Although the concepts learned by an agent are only compromises among

the views of the other agents, the method nevertheless enhances the autonomy of agents using it substantially.

The experimental evaluation shows that the learner agent performs better than or close to teacher agents when

it is tested against the objects from the whole world.

1 INTRODUCTION

In many situations efﬁcient communication is the

main cornerstone of cooperation among agents.

While communication does not always have to in-

volve the use of a language, for many purposes utiliz-

ing a language is a very convenience way to convey

information between agents. In addition to this com-

mon language, a common semantics is also necessary

for communicative agents to interact and understand

each other. Ontology research community tries to ad-

dress issues arisen from violation or relaxation of any

of the above two requirements (Gruber, 1991).

For the sake of simplicity and/or convenience

many researchers in their works have assumed that

it is possible to establish a common language among

agents (e.g., using several variations of agent commu-

nication languages, ACL), and also the agents are pro-

vided with a complete common understanding of all

the concepts they need (e.g., having a common con-

ceptualization). In case that heterogeneity or interop-

erability is a requirement, many researchers assume

that it is possible to have an already existing common

ontology for the agents and that the agent developers

can use this common ontology when designing their

agents, perhaps by calling an ontology service, thus

allowing for easy communication and understanding

among the agents. However, the assumption of ex-

istence of a common ontology is often too strong or

unrealistic. For many application domains, there is

no agreement on ontology for the domain among de-

velopers. Also for many areas existing ontologies are

large, unwieldy and encompass more than what a par-

ticular agent will ever need.

A recent approach is to let the agent have their in-

dividualized ontologies and provide them with learn-

ing mechanisms to learn the concepts they need dur-

ing communication (Williams, 2004) (Steels, 1998).

In our previous work, we have devised a methodology

for having agents learn concepts from several peer

agents (M. Afsharchi and Denzinger, 2006). The ba-

sic idea behind our method is to have an agent, that

realizes that there seems to be a concept it does not

currently know of but expects to need to know, so

queries the other agents about this concept by provid-

ing features (and their values) or examples the agent

thinks are associated with the concept. The queried

agents provide the agent with positive and negative

examples from their understanding of their concepts

(i.e. concepts known by them) which seem to ﬁt the

query which allows the learning agent to use one of

the known concept learning techniques from machine

learning to learn the concept. To help focus on the

negative examples, the teaching agents make use of

526

Safari L., Afsharchi M. and Far B. (2009).

CONCEPTS IN ACTION: PERFORMANCE STUDY OF AGENTS LEARNING ONTOLOGY CONCEPTS FROM PEER AGENTS.

In Proceedings of the International Conference on Agents and Artiﬁcial Intelligence, pages 526-532

DOI: 10.5220/0001662605260532

 SciTePress

selected relations between concepts from their ontolo-

gies. The agent that wants to learn the concept deals

with the fact that the other agents in this group might

not totally agree on which examples ﬁt the concept

and which not, by letting the teaching agents vote on

the examples for which it got contradictory informa-

tion in the ﬁrst place.

The work by Williams (Williams, 2004) intro-

duced the idea of using learning to improve the

mutual understanding about a concept between two

agents. In contrast to our method, Williams agents

uses only a ﬂat repository of concepts and are not in-

volved in a multi agent learning .(Sandip. Sen, 2002)

presents a method of how one agent can train another

agent to recognize a concept by providing selected

positive training examples. The multi-agent dimen-

sion is not addressed and no usage of ontologies is

made. Steels (Steels, 1998), like us, allows for dif-

ferences in ontologies of agents and wants to min-

imize these differences for concepts that are of in-

terest to some of his agents. In contrast to us, his

agents do not use learning to allow for teaching a

concept to an agent. (J. v. Diggelen et al., 2006)

developed ANEMONE which is an minimal ontol-

ogy negotiation environment. A very important point

about ANEMONE is that it is not a concept learn-

ing environment, rather it is an environment which

facilitates ontology mapping between two agents in

a minimal way using a layered approach. Like other

related works, ANEMONE does not talk about fea-

ture diversity and it is not a multi-agent collaboration.

Apart from these works which are directly related to

our work, there are many others in different disci-

plines(e.g. Semantic Web) which we have extensively

reviewed in(M.Afsharchi, 2007).

In this paper, as an extension to(M. Afsharchi and

Denzinger, 2006), we present a comprehensive per-

formance study of the learner agent. We compare

the classiﬁcation accuracy of the learner with teacher

agents against the set of all objects in the agents’

world. The experiments show that the learner perform

better than or close to the teachers in a multi-agent en-

vironment.

The structure of this paper is as follows: in Sec-

tion 2 we give deﬁnitions for the concepts that we use

throughout this paper. In Section 3 the concept learn-

ing mechanism is reviewed, Section 4 introduces our

experimental domain and is followed by our experi-

mental results.

2 BASIC DEFINITIONS

In this section, we provide a brief deﬁnition of each of

the two basic concepts involved in our system which

are ontologies and agents. Also we provide the in-

stantiations of these concepts that we require for our

methods.

2.1 Ontologies and Concepts

A formal deﬁnition for ontology has been presented

in (Stumme, 2002) in which an ontology has been de-

ﬁned as a structure O := (C, ≤

, R, σ, ≤

). C and R

are two disjoint sets with members of C being called

concept identiﬁers and members of R are relation

identiﬁers. ≤

is a partial order on C called concept

hierarchy or taxonomy and ≤

is a partial order on R,

named relation hierarchy.

σ : R → C

is a function providing the argument

concepts for a relation such that |σ(r

)| = |σ(r

)| for

every r

, r

∈ R with r

≤

and for every projec-

tion π

(1 ≤ i ≤ |σ(r

)|) of the vectors σ(r

) and σ(r

)

we have π

(σ(r

)) ≤

(σ(r

)). If c

≤

for c

∈ C, then c

is called a subconcept of c

and c

is a

superconcept of c

. Obviously, the relation ≤

is sup-

posed to be connected with how concepts are deﬁned.

In the literature, taxonomies are often built using the

subset relation, i.e. we have

≤

iff for all o ∈ c

we have o ∈ c

This deﬁnition of ≤

produces a partial order on C

as deﬁned above and we will use this deﬁnition in the

following for the ontologies that our agents use.

Concepts often are seen as collections of objects

that share certain feature instantiations. In this work,

for an ontology O we assume that we have a set of

features F = { f

, ..., f

} and for each feature f

have its domain D

= {v

, ..., v

} that deﬁnes the

possible values the feature can have. Then an ob-

ject o = ([ f

= v

], ..., [ f

= v

]) is characterized by

its values for each of the features (often one feature is

the identifying name of an object and then each object

has a unique feature combination). By U we denote

the set of all (possible) objects. In machine learning,

often every subset of U is considered as a concept. In

this work we want to be able to characterize a concept

by using feature values. Therefore, a symbolic con-

cept c

is denoted by c

([ f

= V

], ..., [ f

= V

]) where

= {v

′

, ..., v

′

} ⊆ D

(if V

= D

then we often omit

the entry for f

). An object o = ([ f

= v

], ..., [ f

]) is covered by a concept c

, if for all i we have

∈ V

. In an ontology according to the deﬁnition

above, we assign a concept identiﬁer to each symbolic

concept that we want to represent in our ontology.

CONCEPTS IN ACTION: PERFORMANCE STUDY OF AGENTS LEARNING ONTOLOGY CONCEPTS FROM

PEER AGENTS

527

2.2 Agents

A general deﬁnition that can be instantiated to most

of the views of agents in literature sees an agent A g

as a quadruple A g = (Sit,Act,Dat, f

). Sit is a set of

situations the agent can be in, the representation of a

situation naturally depending on the agent’s sensory

capabilities, Act is the set of actions that A g can per-

form and Dat is the set of possible values that A g’s

internal data areas can have. In order to determine

its next action, A g uses f

: Sit × Dat → Act applied

to the current situation and the current values of its

internal data areas.

As we want to focus on the knowledge represen-

tation used by agents and how this is used for com-

munication, we have to look more closely at Dat. We

assume that every element of Dat of an agent A g con-

tains an ontology area O

that represents the agent’s

view and knowledge of concepts. There might be ad-

ditional data, beyond features, that the agent requires

from time to time, about concepts and this data is nat-

urally also represented in Dat. Also, there will be

additional data areas representing information about

the agent itself, knowledge about other agents and the

world that the designer of the agent may want to be

represented differently than in O

3 LEARNING CONCEPTS FROM

SEVERAL TEACHERS

In this section we provide a brief description of the

multi-agent concept learning we presented in (M. Af-

sharchi and Denzinger, 2006). As already stated, we

have developed a method that demonstrates how an

agent can learn new concepts for its ontology with the

help of several other agents. This assumes that not all

agents have the same ontology. We additionally as-

sume that there are only some base features F

base

⊆ F

that are known and can be recognized by all agents

and that there are only some base symbolic concepts

base

that are known to all agents by name, their fea-

ture values for the base features and the objects that

are covered by them. Outside of this common knowl-

edge, individual agents may come with additional fea-

tures they can recognize and additional concepts they

know. Given this setting, agents will develop prob-

lems in working together, since the common grounds

for communication are not always there. To come up

with a solution for this problem, agents need to ac-

quire the concepts outside of C

base

that other agents

have, at least those concepts that are needed to estab-

lish the necessary communication to work together on

a given task. The basic idea is to have an agent learn a

required concept (or at least a good approximation of

it) with the help of the other agents acting as teachers.

3.1 The General Interaction Scheme

In the following, A g

refers to the agent that wants to

learn a newconcept and the other agents, A g

,...,A g

will be its teachers. A g

has an ontology O

,≤

,σ

,≤

) and knows a set of features F

Analogously, A g

has as ontology O

= (C

,≤

, ≤

) and knows a set of features F

. For a con-

cept c known to the agent A g

, this agent has in its

data areas a set pex

⊆ U of positive examples for

c that it can use to teach c to another agent. Parts

of Act

are actions

QueryConcept

AskClassify

Learn

, and

Integrate

, while part of the Act

s are the

actions

FindConcept

CreateNegEx

ReplyQuery

ClassifyEx

and

ReplyClass

. For teaching A g

a new concept c

goal

, we have as general interaction

scheme:

After becoming aware that there is a concept that

it needs to learn, A g

performs the action:

QueryConcept

(identiﬁer,{[ f

′

= V

′

],...,[ f

′

]},O

goal

The three parameters of

QueryConcept

allow for

three different ways to identify to the teachers what

A g

is interested in. The parameter identiﬁer allows

A g

to refer to a concept name it observed from other

agents, which means that identiﬁer is an element of

for some agent(s) A g

. By {[ f

′

= V

′

],...,[ f

′

= V

′

]},

A g

can use a selection of features f

′

∈ F

base

and the

valuesV

′

⊆ D

′

that A g

thinks are related to the con-

cept c

goal

. Finally, O

goal

⊆ U is a set of objects that

A g

thinks are covered by c

goal

Each A g

then reacts to A g

’s query by perform-

ing:

FindConcept

(identiﬁer,{[ f

′

= V

′

],...,[ f

′

= V

′

]},

goal

Naturally, each of the parameters can already point to

different concepts that an agent A g

knows of. In fact,

if A g

provides several objects in O

goal

, they might

be classiﬁed by A g

into several of its concepts. So,

A g

ﬁrst collects all the concepts that fulﬁll the query

into a candidate set C

cand

and then it has to evaluate

all these concepts to determine the concept that is the

best ﬁt. So, the output of

FindConcept

is a set of can-

didate concepts C

cand

. To select the “best” candidate

out of C

cand

, there are many different ways how an

evaluation of the candidates can be performed. Each

of the 3 query parts can contribute to a measure that

deﬁnes what is “best”, but how these contributionsare

combined can be realized differently.

ICAART 2009 - International Conference on Agents and Artificial Intelligence

528

The number of examples communicated to A g

by each agent is a parameter of our system. So, in the

next step, each teacher selects the given number of el-

ements out of the set of positive examples, pex

, for

the best candidate concept c

and we call this set p

Again, there are many possible ways how this selec-

tion process can be done, so far we used random sam-

pling of pex

. By then performing

CreateNegEx

the teacher agents produce a given number of (good)

negative examples for c

, which produces the set n

Since every concept c

other than c

(and its subcon-

cepts) can be categorized as a counter concept, the

number of objects associated with these c

s (which

naturally are negative examples) is often very high.

We used both taxonomy information (siblings of

) and a relation

is-similar-to

(similar concepts to

) to select the concepts from which we randomly

selected examples as elements for n

ReplyQuery

) is the last action performed

by a teacher agent before the initiative goes back to

the learner. It sends the result back to the learner.

A g

collects the answers (c

) from all teachers,

pex

goal

= ∪

i=1

and nex

goal

= ∪

i=1

, and then uses

a concept learner to learn c

goal

from the combined ex-

amples (action

Learn

((p

),...,(p

))). Naturally,

the concept learner only uses features and their values

from F

Learning from a group of agents is a very con-

ﬂict prone process compared to just learning from one

agent. It can easily happen that the best concepts c

and c

that A g

and A g

identiﬁed are not the same.

The worst case can be that an example that A g

sent

as being positive for c

goal

A g

sent as a negative one.

But we can also have more indirect conﬂicts where

a learning algorithm simply cannot come up with a

concept description that covers all objects in pex

goal

while not including any objects in nex

goal

. There are

several methods how we can solve this problem and

these methods represent different degrees of willing-

ness to satisfy the teacher agents (by A g

For our system, we have chosen the following

conﬂict resolution to produce c

goal

for A g

. After the

learning component of A g

has performed

Learn

and

produced a more precise c

goal

, A g

will test all ele-

ments of pex

goal

and nex

goal

for correct classiﬁcation

by this new c

goal

. For all the example objects that

are not correctly classiﬁed, we go back to the teacher

agents and ask them to classify these examples ac-

cording to the c

they used to produce their examples.

We then treat the answers as votes and include all pos-

itive examples for which a majority of the teachers

voted, while requiring the exclusion of all negative

examples for which a majority voted. This produces

some kind of compromise concept that might appeal

to most of the teachers (although it might not be iden-

tical to any of the c

The result of this learning/teaching scheme is

the description of c

goal

in terms of A g

’s feature

set F

and an updated ontology O

new

= (C

new

, ≤

, R

, σ

, ≤

4 EXPERIMENTAL RESULTS

To study the performance of the learner agent, we

have chosen the course catalog ontology domain (see

(Il0, )). In the following, we will ﬁrst introduce

this domain and then based on our basic setup from

(M. Afsharchi and Denzinger, 2006) we manage our

learner agent to learn 3 different concepts. Then we

evaluate the performance of the learner agent regard-

ing these newly learned concepts.

4.1 The University Units and Courses

Domain

The university units and courses domain consists of

ﬁles describing the courses offered by Cornell Uni-

versity, the University of Washington and the Univer-

sity of Michigan, together with ontologies for each of

the three universities describing their organizational

structure (see (Il0, ) and (Um0, )). In our exam-

ples, each of the three universities is represented by

an agent (A g

, A g

) and these agents are act-

ing on the one hand side as the teachers to an agent

A g

and on the other side as the agents A g

commu-

nicates with afterwards.

The objects of this domain are the course ﬁles that

consist of a course identiﬁer, a plain text course de-

scription and the prerequisites of a course. All in all,

there are 19061 courses among the three universities

and each university’s ontology has at least 166 con-

cepts on top of their courses. The information for

the three universities does not come with the appro-

priate feature values (at least not directly), instead

we only know the taxonomy and for each unit what

courses belong to it. In order to create the ontologies

, O

and O

for the agents A g

, A g

and A g

we used the learning method of (M. Sahami, 1997)

to create feature-based descriptions of each concept

(as described in Section 2.1). Naturally, the examples

for the learning algorithm were only taken from the

courses of the particular universityand we used all the

courses as examples to achieve a perfect ﬁt. To cover

exactly the courses of a particular unit, we also had

to adjust the initial key word sets of the agents, but

we were able to keep these sets different between the

CONCEPTS IN ACTION: PERFORMANCE STUDY OF AGENTS LEARNING ONTOLOGY CONCEPTS FROM

PEER AGENTS

529

agents. Also, since no examples from other universi-

ties were used for a particular agent, this agent really

reﬂects just the view of this university, so that dif-

ferences in the understanding of a concept (i.e. what

should be taught by what unit) between the universi-

ties are preserved.

4.2 Concepts in Action

It is very important to assess the performance of the

learner agent regarding a newly learned concept. As

stated before our main concern in this paper is to

have agents learn concepts to improve communica-

tion. Needless to say, to communicate about a con-

cept, an agent must distinguish an instance of the

concept (i.e.an object) from other instances. Based

on this fact, we conducted an experiment to see how

A g

classiﬁes objects in U (i.e the set of all possi-

ble objects) when it learns a new concept. Also we

have chosen the set of examples which the majority

of agents vote for, to be learned by A g

First we enabled A g

to learn three different con-

cepts

Greek

Computer Science

and

Mathematics

We allowed A g

to use the most popular way of rely-

ing on a group decision which is to follow the major-

ity of votes as the representative examples of the con-

cept. Then we trained the learner using different per-

centages of positive examples in this area (i.e. Table 1

n% column). These percentages show us the classiﬁ-

cation accuracy of A g

when the learner does not uti-

lize the maximum number of available examples from

the teachers. That is the case when due to the commu-

nication cost, the teachers could not send every possi-

ble example that they possess to the learner. Table 1

shows the classiﬁcation results of the learner for the

three different concepts. In fact this table shows the

number of correctly classiﬁed examples both for pos-

itive examples and negative examples in two separate

columns. We should mention that, when we consider

the examples that a majority of agents agreed upon

as the boundary for the concept in the learner, every

other examples will be tested as the negative exam-

ples by A g

in the testing process. For instance and

for concept

Mathematics

, the majority set has 501

positive examples and the other objects (i.e 19061-

501=18560) could be considered as negative exam-

ples for it.

One interesting preliminary result, that in fact we

expected, was the signiﬁcant increase of correctly

classiﬁed examples when the concept is mostly unan-

imous. For example the programs

Mathematics

and

Greek

have more common courses than

Computer

Science

among three different universities (which

also is very true among other universities). As Ta-

ble 1 shows the accuracy result for

Mathematics

much better than

Computer Science

. The last row

of Table 1 shows the performance of the learner when

it is trained by the whole set of examples it pos-

sess for each concept. For instance, the second and

third columns show that A g

classiﬁed 497 positive

and 17324 negative examples out of 501 positive and

18560 negative examples respectively. Therefore A g

classiﬁed 93% ((497+17324)/(501+18560)) of ob-

jects correctly for

Mathematics

while this accuracy

is 81% ((429+15104)/(505+18556)) for

Computer

Science

and 90% ((170+16991)/(171+18890)) for

Greek

. There is a small “dip” in

Greek

when A g

trained by 70% of examples when the accuracy jumps

to 91% and then comes back to 90%. Despite this

”dip” the learner shows a consistent behavior clas-

sifying positive examples. We conclude that having

agents with close viewpoints helps the learner to have

a concrete understanding of a concept which naturally

leads to a learner with better performance.

To compare the performance of A g

with the

teacher agents we had to compare the classiﬁcation

capability of A g

with A g

, A g

, and A g

respec-

tively. As we discussed earlier, we assume that the

teacher agents have learned the concepts in their on-

tology before they start to teach a concept to the

learner. This learning has been achieved using some

supervised inductive learning mechanisms and using

the example objects that in each agent are associated

with every concept in its ontology. Therefore we are

supposed to simply compare the classiﬁcation efﬁ-

ciency of A g

with A g

, A g

, and A g

Nevertheless we can not guarantee that A g

learns

a concept using the same number of examples as each

teacher agent and ,obviously,the more examples are

provided to the agent the better a classiﬁer it can

learn. This possibility causes an unbalanced situa-

tion in which A g

and other agents can not be com-

pared. To overcome this problem, we have to prepare

a fair situation in which the learner agent classiﬁca-

tion efﬁciency could be compared with each teacher

agent. Therefore we selected a fragment of positive

examples in A g

which is quantitatively equal to the

number of positive examples in each teacher agent to

train A g

with the same number of examples that the

teacher agents utilized to learn the concept before.

Table 2, 3 and 4 show the results of comparisons

of A g

with A g

, A g

and A g

respectively. The

second column in each table shows the number of cor-

rectly classiﬁed examples, both positive and negative,

out of 19061 test examples (i.e. objects in U ) by

A g

. The third column shows the number of cor-

rectly classiﬁed examples by the teacher agent and

ﬁnally the forth row shows the percentage of exam-

ICAART 2009 - International Conference on Agents and Artificial Intelligence

530

Table 1: Correctly classiﬁed examples for concepts

Mathematics

Computer Science

and

Greek

n% Mathematics Computer Science Greek

Positive

Out of

501

Negative

Out of

18560

% Positive

Out of

505

Negative

Out of

18556

% Positive

Out of

171

Negative

Out of

18890

10 400 13125 70 324 11375 61 128 13906 74

20 458 14157 77 339 11934 64 134 14551 76

30 460 15290 82 346 12307 66 142 14958 79

40 472 15267 82 354 12494 67 154 15681 83

50 481 15358 83 367 13053 70 159 16254 86

60 485 15501 83 380 13426 72 163 16808 89

70 484 15691 84 391 13613 73 165 17193 91

80 489 15886 85 399 14172 76 167 17005 90

90 495 16765 90 423 14731 79 170 16995 90

100 497 17324 93 429 15104 81 170 16991 90

Table 2: Comparison of the performance of A g

and A g

concepts correctly

classiﬁed

examples

by A g

correctly

classiﬁed

examples

by A g

ex-

am-

ple

from

A g

Mathematics 15859 83.2% 15886 83.3% 53%

Computer

Science

12962 68.0% 11009 57.7% 43%

Greek 17011 89.2% 16719 87.7% 68%

Table 3: Comparision of the performance of A g

and A g

concepts correctly

classiﬁed

examples

by A g

correctly

classiﬁed

examples

by A g

ex-

am-

ple

from

A g

Mathematics 15780 82.7% 15756 82.6% 44%

Computer

Science

12533 65.7% 11788 61.8% 28%

Greek 15492 81.2% 15121 79.3% 35%

ples that A g

has been trained with, to produce this

result. For instance the ﬁrst row of Table 2 shows

that A g

has correctly classiﬁed 15859 examples out

of 19061 when it is trained by 53% of the whole set

of its positive examples for

Mathematics

. The third

column indicates that 15886 example objects are cor-

rectly classiﬁed by A g

. The last column indicates

that the number of associated examples with concept

Mathematics

in A g

is 53% of examples in A g

Table 4: Comparision of the performance of A g

and A g

concepts correctly

classiﬁed

examples

by A g

correctly

classiﬁed

examples

by A g

ex-

am-

ple

from

A g

Mathematics 15163 79.5% 15086 79.1% 26%

Computer

Science

12805 67.1% 12112 63.5% 39%

Greek 14894 78.1% 14597 76.5% 24%

A very interesting outcome of this experiment is that

A g

in the most cases has a better performance than

the teachers regarding to the learned concept.This em-

phasizes on the fact that A g

learns the compromise

concept and its learning reﬂects a mutual viewpoint of

agents. Therefore it will perform better when it tests

against the objects from the whole world.

As an example concept, if we look at the

Computer Science

we see that A g

is doing better

compared to the other agents. For instance its ac-

curacy is 68% (12962/19061) when it is trained by

43% of the training examples(see table 2). Clearly

this is a better performance than A g

which has clas-

siﬁed 57.7% (11009/19061) correctly. Here we see

that A g

classiﬁes 10.3% better than A g

while this

margin is 3.9% for A g

and 3.6% for A g

. We be-

lieve that this is because the viewpoint of the learner is

closer to A g

and A g

and the compromise concept

in A g

does not have too much of the characteristics

Computer Science

from A g

. Therefore A g

doing much better in classifying objects from U .

The story is different for

Mathematics

. The per-

formance of the A g

is worse than A g

(i.e. 83.2% vs

CONCEPTS IN ACTION: PERFORMANCE STUDY OF AGENTS LEARNING ONTOLOGY CONCEPTS FROM

PEER AGENTS

531

83.3%) but it is better than A g

(i.e. 82.7% vs 82.6%)

and A g

(i.e. 79.5% vs 79.1%) . This observation

shows that the performance of the learner is close to

the performance of other agents and that is because

Mathematics

is more unanimous than

Computer

Science

which makes the viewpoints close to each

other.

REFERENCES

Illinois semantic integration archive.

http://anhai.cs.uiuc.edu/archive/. As seen on Sep

20, 2006.

University of michigan academic units.

http://www.umich.edu/units.php. As seen on

Sep 20, 2006.

Gruber, T. R. (1991). The role of common ontology in

achieving sharable, reusable knowledge bases. In Pro-

ceedings of the Second Iniernanonal Conference on

Principles of Knowledge Representation and Reason-

ing, pages 601–602.

J. v. Diggelen, R. B., Dignum, F., van Eijk, R. M., and

Meyer, J. (2006). ANEMONE: An effective minimal

ontology negotiation environment. pages 899–906.

M. Afsharchi, B. F. and Denzinger, J. (2006). Ontology-

guided learning to improve communication between

groups of agents. In Proceedings of the Fifth Interna-

tional Conference on Autonomous Agents and Multi-

agent Systems (AAMAS06), pages 923–930.

M. Sahami, D. K. (1997). Hierarchically classifying docu-

ments using very fewwords. In Proceedings of ICML-

97, pages 170–178.

M.Afsharchi (2007). Ontology guided collaborative con-

cept learning in multi-agent systems. PhD Disser-

tation, Department of Electrical and Computer Engi-

neering, University of Calgary.

Sandip. Sen, P. (2002). Sharing a concept. In (AAAI Tech

Report SS-02-02).

Steels, L. (1998). The Origins of Ontologies and com-

munication conventions in multi-agent systems. Au-

tonomous Agents and Multi-Agent Systems, 1(2):169–

194.

Stumme, G. (2002). Using ontologies and formal con-

cept analysis for organizing business knowledge. In

Wissensmanagement mit Referenzmodellen - Konzepte

fr die Anwendungssystem- und Organisationsgestal-

tung, pages 163–174.

Williams, A. B. (2004). Learning to share meaning in a

multi-agent system. Autonomous Agents and Multi-

Agent Systems, 8(2):165–193.

ICAART 2009 - International Conference on Agents and Artificial Intelligence

532