Semantic Agent Model for Distributed Knowledge Management.

Abstract

At the beginning of the decade, the Agent Mediated Knowledge Management workshops series [23, 1, 13] as well as Bonifacio's theoretical approach [3] layed the foundations of a new field of distributed knowledge management based upon the agent paradigm. The agent based approach enables key features for knowledge management. The local management of knowledge by agents allows going beyond the limitations of centralized knowledge management. Thus, knowledge can be maintained in each agent at a coarse-grained level, with different representations. Interactions between agents permit us to take into account the social aspect of knowledge and are well suited to represented organizational memories [9]. Van elst also outlined the importance of pro- activeness for knowledge management [22]. In the mean time, the publishing of the agent roadmap in 2003 [17] pointed out the lack of connection between multi-agent systems and semantic web technolo-gies. Since then, many applications and frameworks have been developed to bridge this gap. Semantic web languages and tools are now widely used to repre-sent agents' knowledge. TAGA [25] uses OWL and RDF as knowledge represen-tation in the field of a trading agent competition, using a FIPA compliant frame-work. AgentOWL [15] extends JADE agents with OWL support for their knowledge Base (KB). It also introduces an OWL based semantic agent model. Knowledge agents, introduced by [2], are used for domain specific web search. In this application, agents' KBs are represented in RDF. RDF is also used in CORESE [8] which is a semantic web search engine for corporate knowledge de-veloped within the COMMA (Corporate Memory Management through Agents)

Full text

Page 1

Semantic Agent for Distributed Knowledge
Management
Julien Subercaze and Pierre Maret1
Introduction and motivation

At the beginning of the decade, the Agent Mediated Knowledge Management
workshops series [23, 1, 13] as well as Bonifacio's theoretical approach [3] layed
the foundations of a new field of distributed knowledge management based upon
the agent paradigm. The agent based approach enables key features for knowledge
management. The local management of knowledge by agents allows going beyond
the limitations of centralized knowledge management. Thus, knowledge can be
maintained in each agent at a coarse-grained level, with different representations.
Interactions between agents permit us to take into account the social aspect of
knowledge and are well suited to represented organizational memories [9]. Van
elst also outlined the importance of pro- activeness for knowledge management
[22].
In the mean time, the publishing of the agent roadmap in 2003 [17] pointed out
the lack of connection between multi-agent systems and semantic web technolo-
gies. Since then, many applications and frameworks have been developed to
bridge this gap. Semantic web languages and tools are now widely used to repre-
sent agents' knowledge. TAGA [25] uses OWL and RDF as knowledge represen-
tation in the field of a trading agent competition, using a FIPA compliant frame-
work. AgentOWL [15] extends JADE agents with OWL support for their
knowledge Base (KB). It also introduces an OWL based semantic agent model.
Knowledge agents, introduced by [2], are used for domain specific web search. In
this application, agents' KBs are represented in RDF. RDF is also used in
CORESE [8] which is a semantic web search engine for corporate knowledge de-
veloped within the COMMA (Corporate Memory Management through Agents)

1 Julien Subercaze
Laboratoire Hubert Curien, Saint-Etienne F-42000, France e-mail: julien.subercaze@univ-st-
etienne.fr
Pierre Maret
Laboratoire Hubert Curien, Saint-Etienne F-42000, France e-mail: pierre.maret@univ-st-
etienne.fr
hal-00573409, version 1 -

Page 2

2
European IST project. The JADE framework, which is currently the most used in
research and industry supports natively RDF representing agents' knowledge.
These examples show us that semantic web technologies are widely used for rep-
resenting agent knowledge, and that we can clearly state that the connection be-
tween agent-based knowledge management and semantic web has been made.
However, this example presents the use of semantic for the representation of
agent's knowledge and not for the dynamic part of the agent: its behavior.
Katasonov proposed a Semantic Agent Programming Language (S-APL) [14],
based on BDI reasoning, in which agent behavior are semantically described. Be-
haviors remain programmed in JAVA but are described in RDF syntax. A closed-
world reasoner (CWM2) is used for BDI support. S-APL has three main draw-
backs. First the reasoning is restricted to the closed world assumption whereas
semantic web languages such as OWL make the open world assumption. Secondly
each new function in S-APL has to be programmed in JAVA, which reduces the
interoperability between agents. An agent having some new functions will not be
able to transfer its behavior to another agent since the latter does not own the
JAVA code implementing the functions. Using this approach, agent behavior pro-
gramming is not taking advantage of the semantic web technologies and limits in-
teroperability.
Our motivation is to program agent behavior using semantic web standards us-
ing a finite number of actions that will be used to build complex behaviors. This
language should not refer explicitly to a lower level language and should support
open world reasoning. The latest advances in the field of semantic web have en-
abled rule languages supporting open world reasoning. We base our approach on
the use of semantic rule language to program semantic agents. We aim at design-
ing agents having a knowledge base and behavior base represented using the same
syntax. The use of OWL for knowledge base and semantic rules for behavior al-
lows this feature. In the next section we first discuss the choice of the semantic
rule language that will be used to program agents and then present the design of an
agent programming language and the resulting agent architecture. In section 1.3
we describe the resulting ontological agent model. Section 1.4 shows a practical
example of a SAM behavior and details the different steps of its execution. Sec-
tion 1.5 concerns the implementation of the SAM prototype. We discuss in section
1.6 the perspectives of application of semantic agents in distributed knowledge
management. Our conclusions are presented in section 1.7.


2 http://www.w3.org/2000/10/swap/doc/cwm.html
hal-00573409, version 1 -

Page 3

3

Building agents with semantic rules

Semantic rules are an important part of the Semantic Web project. Figure 1.1
shows the current status of specification in the Semantic Web layer cake. The
logic part, which is of primary interest for us, is still work in progress. This part
describes the semantic rule languages. Currently, for this layer, two proposals are
pending. The most well known one is the Semantic Web Rule Language (SWRL)3
[12]; it is based on a combination of the OWL-DL with the RuleML language.
The second proposal is the Web Rule Language (WRL4) initiative that was in in-
fluenced by the Web Service Modeling Language WSML. Whereas WRL is at a
draft stage, the semantic web community is focusing its research towards SWRL.
Indeed, software such as Protégé5, Pellet6 and Jess7 already provide support for
SWRL. Due to these advances in implementation, it is now possible to develop
agents based on semantic rules. Thus our choice naturally went to SWRL for the
design of the Semantic Agent Model (SAM). SWRL presents two main advan-
tages compared to other rule languages. First because it is an OWL based lan-
guage, it allows writing the rules in terms of OWL concepts (i.e. classes, individu-
als, properties and data values). To these OWL concepts, the SWRL specification
adds several built-ins functions for comparisons, math, strings and time [12]. From
a more agent programming point of view, concepts of the agent knowledge base
can be directly treated in the rule language without loss of expressivity.

3 http://www.w3.org/Submission/SWRL/
4 http://www.w3.org/Submission/WRL/
5 http://protege.cim3.net/cgi-bin/wiki.pl?SWRLTab
6 http://clarkparsia.com/pellet/
7 http://herzberg.ca.sandia.gov/jess/
hal-00573409, version 1 -

Page 4

4

Fig 1.1 TheSemantic Web Layer Cake

The second benefit of SWRL resides in its logical foundations. SWRL com-
bines OWL-DL (decidable version of OWL) with Rule Markup Language
(RuleML). Thus, it can be roughly considered as the union of Horn-Logic and
OWL based on the description logic SHOIN. Consequently the expressivity of
SWRL comes at the price of decidability [19]. SWRL is not decidable (SWRL
full). However, a subset called DL Safe SWRL rules is decidable. For the agent
development and for reasoning, DL Safe SWRL rules are more expressive com-
pared to other rule languages. Indeed, most of the rule-based agents are based on
Prolog supporting Horn Clauses. Researches are currently going on for imple-
menting DL reasoning in Prolog but none is currently available for agents. Practi-
cal advantages for using Description Logic in the field of Multi-Agent Systems
(MAS) has been shown in [18], especially in the field of information retrieval.

Architecture design

Programming agent behavior using a rule language can be carried out in two
ways. The first way consists in extending a logic programming language in order
to support traditional agent features (i.e. message passing, threading, etc.). The
second way consists in building a layered architecture using the rule language at
an upper layer. Some agent features are delegated to a lower layer. In this type of
architecture, the lower level language (i.e. Java, C++,etc.) is commonly used to
handle communication, file access, thread management, etc. The main idea behind
this approach is to reuse the required features for MAS that are already imple-
mented in another language and to define an agent interpreter to support a particu-
lar architecture, such as BDI for instance. The literature shows examples of both
approaches. Clark et al. [7] follows the first approach by extending Qu-Prolog
with multi-threading support and inter-thread message communication. However,
hal-00573409, version 1 -

Page 5

5
this approach is not scalable and does not comply with the Agen Communication
Language (ACL) as specified by the FIPA8. FIPA-ACL is currently recognized as
the standard for agent communication and ensures interoperability between MAS
frameworks. S-APL that we discussed in the previous section follows the same
approach but some direct calls to JAVA functions are directly inserted into the
rules.

SAM Architecture

Standard MAS languages rely on the second approach. Agent0, the first agent
dedicated language, which is an implementation of Shoham's Agent Oriented Pro-
gramming was developed on top of LISP. Similarly, 3APL, 3APL-m, JA- SON
and the BDI agent system Jadex are based on JAVA. Our architecture follows the
same approach. The specificity of our approach is to rely on a finite number of ac-
tions, in order to ensure interoperability of agents' behaviors.
In short, our approach results in the layered architecture (Fig.1.2) :
1. Knowledge Base
2. Engine
3. MAS Framework and low level actions

8 http://www.fipa.org/repository/aclspecs.html
hal-00573409, version 1 -