AGENT ONTOLOGY INTEROPERABILITY APPROACH
FOR MAS NEGOTIATIONS IN VIRTUAL ENTERPRISES
X. H. Wang, T. N. Wong and G. Wang
Department of Industrial and Manufaturing Systems Engineering, The University of Hong Kong
Pokfulam Road, Hong Kong
Keywords: Negotiation, Ontology Matching, Multi-agent System, Virtual Enterprise.
Abstract: In supply chain management, a Virtual Enterprise (VE) is a dynamic alliance of partner companies. Multi-
agent systems (MAS) have been introduced to facilitate negotiations among VE members. From the
perspective of knowledge management, heterogeneous VE members utilize different knowledge structures
and terminologies in their representative agents. To encourage their collaborative coordination and realize
mutual understanding, agent ontology interoperability should be reached. In this paper, an approach for
semantic ontology matching is proposed to generate correspondences among heterogeneous ontologies
embedded in MAS; additionally, an ontology correspondence generation and negotiation protocol is
developed to realize agent ontology interoperability in MAS negotiations.
1 INTRODUCTION
A Virtual Enterprise (VE) is a dynamic alliance of
partner companies, aiming at seeking more market
opportunities to provide high-quality with low-cost
services and products as fast as possible
(Camarinha-Matos et al, 1999). Generally, a VE is
composed of a VE initiator and several distributed
and heterogeneous VE partners. To respond to a
market opportunity, the VE initiator relies on its
partners to operate, that is, the functioning of a VE
relies on the collaboration and interaction of its
partners. Therefore, it is important for the VE
initiator to select appropriate partners to create the
initial VE.
In the VE formation phase, the initiator
negotiates with its potential partners to reach
agreement on the cooperation issues in order to form
the initial VE. In the subsequent VE functioning
phase, the initiator negotiates with its partners to
contract with each other. With the advance of agent
technology, multi-agent systems (MASs) have been
introduced to provide an effective and efficient
environment for VE formation and operation
(Norman et al, 2004). Within the MAS, agents are
established to be responsible for various functions of
different VE members. However, different VE
members may use different structures and
terminologies to organize and represent their
knowledge; it is important but difficult for agents to
reach mutual understanding through heterogeneous
knowledge representations. In order to achieve
semantic interoperability among agents, many
researchers have proposed to adopt the ontology-
based approach to support MAS negotiations in the
VE (Chen et al, 2008; Lo et al, 2008; Trappey et al,
2009; Garcia-Sanchez et al, 2009). Some are
concentrating on developing ontology matching
mechanisms to reach semantic interoperability
(Chen et al, 2008; Lo et al, 2008); while some are
focusing on realizing MAS automated negotiations
(Trappey et al, 2009; Garcia-Sanchez et al, 2009). In
most of the current applications, there are only a few
publications focusing on an entire framework to
illustrate how agent ontology interoperability is
reached, how semantic matching is conducted, and
how ontology based MAS negotiations is realized.
The objective of this paper is to establish an
ontology-based framework to achieve agent
ontology interoperability and to realize automated
MAS negotiations in the VE.
The remainder of this paper is organized as
follows: Section 2 presents some related work.
Section 3 outlines the MAS framework and semantic
matching approach. Section 4presents system
implementation and experiment evaluation. Section
5 presents conclusions and future work.
149
H. Wang X., N. Wong T. and Wang G. (2010).
AGENT ONTOLOGY INTEROPERABILITY APPROACH FOR MAS NEGOTIATIONS IN VIRTUAL ENTERPRISES.
In Proceedings of the 2nd International Conference on Agents and Artificial Intelligence - Agents, pages 149-154
DOI: 10.5220/0002704301490154
Copyright
c
SciTePress
2 RELATED WORK
2.1 Ontology-based MAS for VE
Many researchers have engaged in providing
solutions in this field to reach mutual understanding
and automated MAS negotiations in VE.
In Lo et al (2008)’s application, the VE initiator
builds up domain ontology with an ontology
vocabulary inside its database, which can provide
flexible matching with different enterprises. Trappey
et al (2009) developed a JADE-based workflow
management system, where the workflow ontology
is constructed to represent relationships between
workflows, resources and actors. Garcia-Sanchez et
al (2009) designed and implemented a JADE-based
platform for the provision of semantic web services,
where ontologies is used to describe application
domain knowledge, agent local knowledge,
negotiation knowledge and semantic web services
knowledge.
The above applications are focused on
developing MAS platforms. Detail information
about utilizing heterogeneous ontologies is not
available. Therefore, a series of semantic ontology
matching approaches are proposed to fill the gap.
2.2 Ontology Matching
Ontology is increasingly considered as an essential
factor for reaching interoperability across
heterogeneous systems. Meanwhile, ontology
matching has been introduced to combine distributed
and heterogeneous ontologies.
Choi et al (2006) conducted a survey on ontology
matching, aiming at providing a comprehensive
understanding of ontology matching. Laera et al
(2007) proposed an argumentation based ontology
matching for MAS. Correspondence matching
repository is constructed, where candidate matchings,
ontology mismatches, matching preference and
candidate threshold are stored in. Bollegala et al
(2007) engaged in conducting query-based semantic
matching studies, where various types of query
ontologies were constructed to enable agents’
learning ontology models.
As revealed in the above publications, previous
research efforts have been mainly spent on the basic
ontology matching approaches. As presented in this
paper, a series of ontology-based semantic matching
approaches are proposed to reach agent ontology
interoperability in VE negotiations.
3 APPROACH
3.1 Ontology-based MAS Framework
An ontology-based MAS framework has been
developed to enable distributed and heterogeneous
VE members to communicate and negotiate. Figure
1 depicts the architecture of the system framework,
which comprises the following elements:
Potential Partners
BA
KMA
CA
SA
CAi
SA
BAi
BAi
CAi
SAi
SAi
SA
BA
BA
<Negotiation>
Market
Opportunity
Knowledge
Repository
Partner KB
Partner KB
Partner KB
Partner KB
TEA
VE Partners
Performance
TDA
1. Individual ontology models
2. Correspondence library
3. Correspondence candidate library
1. Individual ontology models
2. Correspondence library
<Contracting>
Figure 1: Ontology based MAS system architecture.
Buyer Agent (BA): In this study, BA is on behalf
of the VE initiator, which initiates the
preliminary requirements to select its partners.
Seller Agent (SA): SA is on behalf of VE
partners, which responds to BA’s requirements
and negotiate with it to reach agreements.
Task Decomposer Agent (TDA): Receive
preliminary requirements from BA and
decompose them into small sub-tasks.
Coordinator Agent (CA): Receive sub-tasks
from TDA, and sends them to BA and
initializes the negotiation.
Task Evaluator Agent (TEA): Evaluate the
partners’ performances, and record it into VE
initiator’s knowledge repository.
VE Initiator’s Knowledge Repository: Store
individual ontology models, correspondence
libraries, and correspondence candidate
libraries.
VE Partner’s Knowledge Base: Store individual
ontologies, correspondence libraries with
different partners.
Correspondence library: Record historical
correspondences used in past negotiations
between BA and other agents.
Correspondence candidate library: Record
correspondence candidates obtained from
semantic matching process, using domain
individual ontologies.
ICAART 2010 - 2nd International Conference on Agents and Artificial Intelligence
150
3.2 Correspondence Generation
and Negotiation Protocol
Figure 2 depicts ontology based correspondence
generation and negotiation protocol, which aims at
realizing mutual understanding and automated
negotiation among VE members. Specification for
this protocol is provided in Table 1.
Figure 2: Ontology correspondence generation and
negotiation protocol.
Here, BT is a term using buyer’s terminology,
ST
j
is a term using seller’s terminology in form of
(BT, ST
j
) in buyer’s correspondence library; Sagent
is an agent participating in negotiations using (BT,
ST
j
) as a correspondence; TT is a term translated
according to buyer-seller correspondence candidate
library; translated indicates that (BT, TT) is a
correspondence candidate pair, not a correspondence
pair; RT
i
is a term in form of (BT, RT
i
) in other
partner’s correspondence libraries.
Table 1: Interpretation of negotiation messages.
Message
Name
Functionality
INFORM Send a term to the receiver agent;
Ask it to check its correspondence
library using the received
information.
CONFIRM Confirm the correspondence with the
initiator.
NOT_UNDER
STAND
No correspondence is in the partner’s
library.
CALL FOR
BID
Initialize the negotiation with a
preliminary bid.
PROPOSE Respond to CALL FOR BID
message.
NEGOTIATE Negotiate with each other.
CONTRACT Reach agreement and contract.
FAILURE No agreement is made.
3.3 Knowledge Representation
Ontology is used to describe domain knowledge of
individual VE partners. In this study, two types of
individual negotiation ontologies are constructed, i.e.
Buyer ontology and Seller ontology, as shown in
Figure 3.
Figure 3: Partial ontologies for buyer and seller 1.
AGENT ONTOLOGY INTEROPERABILITY APPROACH FOR MAS NEGOTIATIONS IN VIRTUAL ENTERPRISES
151
3.4 Semantic Ontology Matching
The purpose of ontology matching is to find out the
correspondences between two separate ontologies.
By doing so, heterogeneous ontologies can reach
mutual understanding among each other.
Definition 1 (Ontology Matching System):
Suppose O and
'
O
are two ontologies, which are
defined as O = <C, R, I> and
>=<
''''
,, IRCO
. Here,
C stands for concepts, R stands for relations, while I
stands for instances. An ontology matching system
MS
is defined as a triple (
MOO ,,
'
). M is the
correspondence between two ontologies, which is
defined as (
σ
,,
'
ee
). Here,
'
and ee
are concept or
attribute in ontologies O and
'
O
, respectively;
σ
is
similarity value between
'
and ee
.
The semantic of heterogeneous ontologies is
calculated by ontology matching. This study
proposes three types of ontology matching methods,
which are explained in the following sections.
3.4.1 Name-based Term Matching
In name-based matching process, firstly, names of
elements (concept and attribute) should be stemmed
and pre-processed into atomic terms. Secondly,
WordNet is introduced to find semantically similar
terms among heterogeneous ontologies (Cognitive
Science Laboratory, Princeton University, 2006).
Algorithm 1: Term similarity calculation
Input: Term T from ontology O, Term
'
T
from
ontology
'
O
Output: <
σ
,,
'
TT
>, correspondence between T and
'
T
Initialize similarity value
),(
'
TT
σ
= 0;
Get all the senses and their hypernym of T
and
'
T
respectively, i.e., T (*) and
'
T
(*);
Calculate the lengths of all paths between T
and
'
T
, and get the shortest path;
Suppose L is the length of the shortest path
between T and
'
T
, then similarity
),(
'
1
TT
σ
can be
calculated by (1):
L
eTT
α
σ
=),(
'
1
(1)
Calculate the depths of all terms in set of
(*)'(*) TT
.
Suppose H is the biggest depth between T and
'
T
,
then similarity
),(
'
2
TT
σ
can be calculated by (2):
HH
HH
ee
ee
TT
ββ
ββ
σ
+
=),(
'
2
(2)
Final term similarity can be calculated by (3):
),(),(),(
'
22
'
11
'
TTTTTT
σωσωσ
+=
1 where
21
=
+
ω
ω
(3)
3.4.2 Structure and Constraint
based Attribute Matching
Ontology structure and constraint defined within
each element have significant effect on elements’
semantics. For concepts, a child concept will inherit
the semantics of its father concept; for attributes,
different attributes’ data types represent different
semantics. In this study, attributes are main
components which make up of the negotiation
messages, a structure and constraint based hybrid
matching algorithm for attribute is proposed.
Algorithm 2: Attribute similarity calculation
Input: O = <C, R, I> and
>=<
''''
,, IRCO
; Attribute
'
0
r
and its related concept
'
0
c of ontology
'
O
.
Output: <r,
'
0
r ,
max
σ
>, the highest similarity between
all attributes in O and attribute
'
0
r from
'
O
.
Find all attribute r in O of the same data type
with
'
0
r , and their related concept names c.
Calculate similarities between qualified
attribute r and
'
0
r using formula (4):
),(),(
),(),(),(
'
044
'
013
'
012
'
011
'
0
rrrc
rcccrr
σασα
σασασ
+
++=
(4)
Then the highest similarity can be obtained by
formula (5):
max
σ
= max (
),(
'
0
rr
σ
) (5)
3.4.3 Distance between Data Patterns
based Instance Matching
The contents of instances reveal some correlations
among different ontologies. Therefore, an instance
matching is proposed based on auxiliary information,
such as data pattern, value distribution, average, etc.
Algorithm 3: Instance similarity calculation
Input: Instance strings SB
and SS
Output: Distance between input strings, d (SB, SS)
Pre-process the attribute data according to the
following rules: Turn all numerals into symbol
“0”; Turn all alphabets into symbol “1”; Turn
ICAART 2010 - 2nd International Conference on Agents and Artificial Intelligence
152
Table 2: Similarity results for correspondence candidate generation using three different methods (Buyer and Seller 1).
Correspondence pair Attribute method Instance method Combined method
Buyer Seller 1 Similarity Rank1 Distance Rank2 Average
(Rank1, Rank2)
Rank3
resolver resolver 0.4901 1 0 1 1 1
initiator proposer 0.3469 3 0 2 2.5 2
issue name item name 0.6073 2 1 1 1.5 1
issue value item value 0.5944 2 0 1 1.5 1
product product 0.4414 3 0 1 2 1
name name 0.5606 4 0 1 2.5 1
address address 0.5204 2 0 1 1.5 1
issue name item name 0.5947 1 1 1 1 1
status status 0.7506 1 0 1 1 1
payment pattern payment pattern 1 1 8 4 2.5 4
due date value time value 0.6503 2 0 1 1.5 1
warranty value service time 0.3994 4 0 3 3.5 3
quantity value order size 0.6025 3 1 2 2.5 2
price value price value 1 1 0 1 1 1
adjustment adjustment 0.6237 1 0 1 1 1
all “ ,” into symbol “X”; Turn all white space
into symbol “Y”; Turn all “http://” into
“*******”; Turn all “@-:/.” into symbol “#”.
Afterwards, the contents of different instances
are transformed into a series of similar strings.
Introduce Edit Distance to calculate similarity
of strings (Navarro, G., 2001).
Definition 2:
SB stands for a
string from buyer; |SB| stands
for the length of SB;
i
SB
stands for the i
th
character of SB, for an integer i
{1…|SB|};
jiiji
SBSBSBSB ...
1... +
=
stands for a partial string
from
i
SB
, where i > j;
Strings for sellers are defined in the same way.
Definition 3: d(SB, SS) indicates the distance
between two strings SB
and SS, which is the minimal
cost of a sequence of operations that transform SB
into SS. Here, the operation refers to delete, insert or
substitute a character.
Definition 4: Define a matrix C
|SB|, |SS|
, where C
i, j
represents the minimum number of operations
needed to match
i
SB
..1
to
j
SS
..1
. d(SB, SS) is computed
as follows:
C
i, 0
= i; C
0, j
= j;
C
i, j
= if (
i
SB
=
j
SS
) then C
i-1, j-1
else 1 + min (C
i-1, j
, C
i, j-1
, C
i-1, j-1
)
d (SB, SS) = C
|SB|, |SS|
.
4 IMPLEMENTATION
AND EVALUATION
In this study, the MAS is implemented using JADE.
As a simple example in supply chain management, a
company has to purchase a product from two
potential suppliers. The case is therefore to establish
a VE with two of its suppliers. Accordingly, the
MAS comprises three VE members, here the VE
initiator acts as the buyer part.
Since different VE members are independent
companies, it is usual for them to adopt different
terminologies, even though they are describing the
same semantics. It is easy for human experts to
identify the representations, but difficult for agents
to recognize automatically. Therefore, semantic
ontology matching approaches are developed to
reach agent interoperability. As detailed in Section
3.4, the word method is a basic method to calculate
term similarities; the attribute and instance methods
are based on hybrid criterions and data patterns of
instances, respectively.
In this study, suppose that two sellers share a
same ontology structure. Two separate ontologies
are shown in Figure 3. Correspondences are
generated, and performances of different methods
are compared, which are as shown in Table 2.
To evaluate the performance of different
matching methods, four typical evaluators in
Information Retrieval (IR) are adopted (Islam, A.,
2008). In the following, TP stands for True Positive
(how many correspondences were selected with
right meanings); FP stands for False Positive (how
AGENT ONTOLOGY INTEROPERABILITY APPROACH FOR MAS NEGOTIATIONS IN VIRTUAL ENTERPRISES
153
many correspondences were not selected, which are
actually with right meanings); FN stands for False
Negative (how many correspondences were selected
with wrong meanings).
The evaluators are listed as below:
Precision (P): P = TP / (TP + FP)
Recall (R): R = TP / (TP + FN)
F-Measure (F): F = 2PR / (P+R)
Accuracy (A): A = (TP + FP) / (TP + FP + FN)
Figure 4 shows the experiment evaluation results.
Figure 4: Performance of attribute & instance
& combination methods (Rank 1 as threshold).
Figure 4 illustrates the comparison of the
attribute, instance and combination methods, where
rank No. 1 correspondence candidate is adopted as
the threshold. It indicates that with the threshold set,
performance of the attribute method is the worst and
that of the instance method is the best.
However, in reality, the instance method is less
restrained since contents of instances can be readily
modified manually, the performance cannot be very
stable. For this reason, the combination method is
adopted to balance the attribute method and the
instance method in a stable and well-performed way.
5 CONCLUSIONS
This paper presents an ontology-based approach to
achieve agent ontology interoperability in MAS
negotiations in the VE formation process. First of all,
an ontology-based correspondence generation and
negotiation protocol is proposed to provide a way
for agents to interact with each other. Secondly,
three semantic ontology matching methods are
proposed, where the combination method with Rank
No.1 correspondence candidate as threshold is
adopted as the most appropriate method to realize
agent ontology interoperability.
The research is still in progress. Future
enhancements will be developed from two aspects:
Firstly, new knowledge is created in every contract
round. It is required to consider how to update the
current knowledge libraries and to manage VE
knowledge evolution in an effective and efficient
way. Secondly, for different roles of agents, i.e.
buyer agents and seller agents, in order to ensure the
security of the negotiation platform, different
members should be assigned different levels of
authority to access the system. Therefore, a
knowledge access control mechanism is to be
developed in the future.
REFERENCES
Bollegala, D., Matsuo, Y., Ishizuka, M., 2007. Measuring
semantic similarity between words using web search
engines. WWW2007, May 8-12, 2007, Banff, Alberta,
Canada, 757-766.
Camarinha-Matos, L.M., Afsarmanesh, H., (1999). The
virtual enterprises concept: Infrastructures for virtual
enterprises. Kluwer Academic Publishers: Boston,
MA, 3-14.
Chen, Y.J., Chen, Y.M., 2008. Development of a
mechanism for ontology-based product lifecycle
knowledge integration, Expert Systems with
Applications, doi:10.1016/j.eswa.2008.01.049.
Choi, N., Song, I.Y., Han, H., 2006. A survey on ontology
matching. ACM SIGMOD record, 35(3), pp. 34-41.
Cognitive Science Laboratory, Princeton University, 2006.
WordNet a lexical database for the English language,
viewed 10 May 2009. <http://wordnet.princeton.edu/>
Garcia-Sanchez, F., Valencia-Garcia, R., Martinez-Bejar,
R., Fernandez-Breis, J.T., 2009. An ontology,
intelligent agent-based framework for the provision of
semantic web services. Expert Systems with
Applications, 36: 3167-3187.
Islam, A., Inkpen, D., 2008. Semantic text similarity using
corpus-based word similarity and string similarity.
ACM Transactions on Knowledge Discovery from
Data, 2(2): Article 10.
Laera, L., Blacoe, L., Tamma, V., 2007. Argumentation
over ontology correspondences in MAS. In
Proceedings of the Sixth International Joint
Conference on Autonomous Agents and Multi-Agent
Systems (AAMAS 2007), Honolulu, Hawaii, USA.
International Joint Conference on Autonomous Agents
and Multi-Agent Systems (AAMAS), Hawaii, USA.
Honolulu:acm, pp. 1293-1300.
Norman, T.J., Preece, A., Chalmers, S., etc., 2004. Agent-
based formation of virtual organisations. Knowledge-
Based Systems, 17: 103-111.
Lo, W.S., Hong, T.P., Jeng, R., 2008. A framework of E-
SCM multi-agent systems in the fashion industry,
International Journal of Production Economics,
doi:10.1016/j.ijpe.2007.09.010.
Navarro, G., 2001. A guided tour to approximate string
matching. ACM Computing Surveys, 33(1): 31-88.
Trappey, C.V., Trappey, A.J.C., Huang, C.J., Ku, C.C.,
2009. The design of a JADE-based autonomous
workflow management system for collaborative SoC
design. Expert Systems with Applications, 36, 2659-
2669.
ICAART 2010 - 2nd International Conference on Agents and Artificial Intelligence
154