SEKE 2012 Proceedings - Knowledge Systems Institute

the ontology hierarchy; learning a new concept (in the learner
agent); selecting positive and negative examples for a certain
concept (in the teacher agents); searching the ontology
hierarchy for the best matched concept to teach the learner
agent; managing tie strengths between agents; and finding
peers.
B. Ontology
There are several definitions of ontology. We will use
Daconta's definition [2] as it is relevant to our work: “Ontology
defines the common words and concepts (meanings) and their
relationships used to describe and represent an area of
knowledge, and so standardize the meanings.”
If two agents use the same ontology or ar e able to
understand each other’s ontology, communication between
them is p otentially possible. Normally, diverse agents use
different ontologies. In this case they need a mechanism to
understand each other. In this paper, we illustrate how a single
learner agent can learn new concepts from different teacher
agents. Those teacher agents do not need to agree on the same
definition of the new concept; their understanding of this new
concept may be close but slightly different from each other.
C. Social networks
By social networks we do not mean Facebook, Twitter or
other related web services. In this work, a social network is a
set of actors (e.g. human, process, agent, document repository)
and relationships between them. It can be represented as a set
of nodes that have one or more types of relationships (ties)
between them [3]. Using social networks gives us flexibility in
dealing with the concepts in heterogeneous ontologies. It
allows agents to understand the meaning of the same concept
even though its definition might be slightly different in each
agent’s ontology.
The strength of a tie is affected by several factors.
Granovetter [4] prop osed four dimensions that may affect tie
strength: the duration of the relationship; the intimacy between
the two actors participating in the relationship; the intensity of
their communication with each other; and the reciprocal
services they provide to each other. In social networks of
humans, other factors, such as socioeconomic status,
educational level, political affiliation, race and gender are also
considered to affect the strength of ties [5]. Structural factors,
such as network topology and information about social circles,
may also affect the tie strength [6]. Gilbert et al [7] suggest
quantitative measures (variables) for tie strength including
intensity variable, days passed since the last communication
and duration [7]. Another variable that may affect the strength
of the tie is the neighborhood overlap variable [8] which refers
to the number of common friends the two actors have. Petróczi
et al [9] introduced mutual confidence between the actors of
social networks. We proposed in [10] a method to calculate the
strength of ties between agents in a social network using
Hidden Markov Models (HMM) [11]. We showed that tie
strength depends on several factors: Closeness factor: by
measuring how close two agents are to each other (i.e. the
degree of similarity between the two ontologies used by the
two agents participating in the relationship); Time-related
factor: combines all time factors that affect the strength of the
relationship (e.g. duration of the relationship, frequency of
communication between the two agents, time passed since the
last communication); Mutual confidence factor: clarifying the
nature of the relationship under measure, if it is a one-sided
relationship or a mutual relationship. Then we built an HMM
model to measure the strengths of ties between agents in a
social network using those factors.
D. Literature review
In this section, we describe some of the previous work done
in the concept learning area.
In [12], Steels uses a distributed multi-agent system where
no central control is required. He describes a “language game”
in which each agent has to create its own ontology based on its
experience of its environment. The agents communicate with
each other to stabilize to a common ontology or a shared set of
lexicons to be used afterwards in their interactions.
Palmisano [13] uses a m achine-learning approach for
creating a shared ontology among different agents that use
different ontologies. The main purpose of Palmisano’s work is
to make each agent maintain its own ontology and at the same
time keep track of the concepts’ meanings in other agents’
ontologies it communicates with.
Afsharchi tries in his work [14] to enable an agent to learn
new concepts from a group of agents then represent this new
concept in its own terminology without the need of a shared
ontology [15]. He deals with a multi-agent system as a group of
agents with different ontologies and any of them can learn a
new concept by asking the other agent about this concept and
then representing it using its own taxonomies.
III. CONCEPT LEARNING FRAMEWORK
In our framework, we assume that in a society of n agents
Ag 1 , Ag 2 … Ag n , each agent Ag i controls a repository R i
(Figure 1). Each repository uses an ontology (O i ) that consists
of a set of concepts, relationships. Each concept C in each
repository possesses some supporting documents to represent
instances of the concept.
R 1
R 2 R n
Ontology
Documents
...
Ontology
Documents
Ontology
Documents
MAS 1 MAS 2
...
MAS n
Figure 1. System Architecture
The goal of this paper is to advance the concept learning
mechanism based on semantic interoperation between concept
learning and semantic search modules based on MAS proposed
in [16][17]. In this prototype system, each MAS controls a
knowledgebase with a certain ontology. The system is expected
to have two main modules: Concept Learning module [18]; and
Semantic Search module [19]. In this paper we focus only on
advancing the concept learning module by applying the social
262