Sie sind auf Seite 1von 6

An Agent-based Knowledge Management Framework

Martin S. Lacher, Michael Koch


Technical University Munich
Arcisstr. 21
80290 Munchen
Germany
flacher,kochmg@in.tum.de

Abstract One lesson learned from these e orts is, that KM can
We propose a framework for a Knowledge Man- always only be e ective with a holistic view. Bullinger
agement (KM) Information Technology (IT) sys- et al. propose in (Bullinger, Worner, & Prieto 1998)
tem. Our framework can integrate existing solu- information and communication technology, human re-
tions for KM IT support. The framework con- source management and organization as the crucial per-
sists of loosely coupled agent-based components spectives of a holistic view on KM for its implementa-
and can easily be extended with further compo- tion in organizations. More genericly put, KM includes
nents. computer science, psychology/pedagogics and business
We identify core components and functionalities aspects. We will derive the needs for a KM support
of a KM support system. We motivate our pro- system from previous research in KM and include the
posal with a holistic perspective on KM. We derive constraints in computer science for implementation.
a classi cation of knowledge in organizations for The second lesson learned from existing KM IT ef-
which management can and should be supported forts is that patchwork should be avoided. Proprietary
by an IT system. solutions for part aspects of KM create new barriers
We introduce an agent-based KM IT architecture
that includes the derived functionalities. The ar- for the knowledge ow instead of facilitating it. It can
chitecture is structured according to knowledge also be seen that certain components (e.g. ontologies,
classes and group-domains of knowledge. We fur- knowledge domain classi cations) are a crucial basis for
ther propose new implementation schemes and most functionalities in a KM system and thus have to
present existing solutions from research for our be accessible to all components and centralized. Ex-
architecture. Fields for further research in KM IT isting software solutions for IT support in business like
support are identi ed. work ow management systems or document reposito-
Our framework is intended to contribute to a uni- ries must be integratable. A KM IT framework also
ed view of requirements of a KM support system. has to be easily expandable by new components.
We further aim to give an overview of applicable We propose an open and high-level IT support for
and available technologies for implementation. KM. Our approach is an agent-based framework that
includes vital components of an open KM architecture.
Introduction We will derive the central KM components from re-
The need for e ective Knowledge Management (KM) quirements identi ed in KM research in the section
in organizations has been largely recognized. Tools and Core Knowledge Management support functionalities.
strategies from research have been put to work in orga- We will then transfer those requirements into a KM
nizations. However, many of the Knowledge Manage- IT architecture in the section Information Management
ment e orts were futile. There are two main reasons Architecture. The implementation of several core com-
for this: on the one hand, merely the information man- ponents of the architecture will be proposed in the sec-
agement tools were implemented and other aspects of tion Implementation Concepts. We will conclude what
Knowledge Management were left aside. On the other we could contribute in the Conclusion.
hand, the existing information management tools are
proprietary, heterogeneous and thus as a whole do not Core Knowledge Management support
at all achieve their goal: the facilitation of the knowl-
edge ow (Borgho & Pareschi 1998). The knowledge
functionalities
ow describes the distribution, recombination and cre- We are taking a holistic perspective on KM. For the
ation of new knowledge and is identi ed as the key lever- derivation of the requirements of a KM IT support sys-
age to KM in (Borgho & Pareschi 1998). tem we have to take related issues from psychology and
business into account. Our aim is the identi cation of
Copyright c 1999, American Association for Arti cial In- principles for an IT system to facilitate the knowledge
telligence (www.aaai.org). All rights reserved. ow (Borgho & Pareschi 1998) in an organization.
A central notion in KM is the di erentiation between comply, but the term knowledge could be used without
explicit knowledge and tacit knowledge, rst introduced changing the essence of the model.
in (Polanyi 1958). Explicit knowledge can be easily
explicitly coded (e.g. into natural language) into infor-
mation and can be characterized as theoretical knowl-
edge. Tacit knowledge is hard to express and can be
characterized as experience. In (Nonaka & Takeuchi
1995), the transfer process between the di erent kinds
of knowledge is analyzed. They extend the analysis
of Anderson's ACT-theory (Singley & Anderson 1989),
which is limited to the transfer of explicit knowledge
to tacit knowledge. IT systems can always only sup-
port the transfer of explicit knowledge. However, there
is no principal reason why tacit knowledge should not
be explicatable. It is more an economic question, since
explication of tacit knowledge involves considerable ef-
fort. We will assume for our work, that relevant tacit
knowledge has been explicated, and thus can be dealt Information demand
with in an IT system. We consider information to be
explicitly coded knowledge, and knowledge to be infor- Subjective Information requirement
mation that has been internalized by a person and is Information available
functionally applicable. Objective Information requirement
To create a system that facilitates the knowledge Actual Information level
ow, we identify and classify the relevant knowledge in
an organization. We choose an inductive-analytical ap- Figure 1: Information management model in (Heinen
proach for knowledge accounting presented in (North, & Dietel 1991)
Probst, & Romhardt 1998): the intangible assets moni-
tor, rst introduced in (Sveiby 1998). Sveiby proposes, The subjective information requirement is what the
that intangible assets (in contrast to visible equity) in user considers the information relevant to perform the
an organization can be classi ed into external structure task. The objective information requirement is what is
of the organization, internal structure of the organiza- really relevant information to perform the task. The
tion and sta competence. The internal structure is information available is the information the user can
de ned as business processes and technologies, whereas access at the moment. The information behavior of the
the external structure is de ned as relations along both user leads to the user's information demand, which is
directions of the value chain (i.e to customers and sub- only a subset of what the subjective requirement is. The
contractors). Sta competence is seen as functionally intersection of the four sets of information is the actual
applicable knowledge. level of information the user has available to perform
From Sveiby's model we derive three classes of knowl- the task. We identify two ways to increase the infor-
edge that we can capture: explicit technical knowl- mation level with the user. One way is to make the
edge (from internal structure and sta competence), user more aware of the information available, and thus
tacit process knowledge or process experience enlarge the subjective information requirement. The
(from internal structure and sta competence) and re- other way is to supply information that has not been
lational network knowledge (from external struc- demanded, but is assumed to be of relevance. We do
ture). not think however, that the third way, determining the
We claim that knowledge necessary for the perfor- objective information requirement for a task, is feasible
mance of a task has to be provided to the user pro- for an IT system.
actively. This claim is based on research from so- There are several other advantages to pro-active
cial psychology (Hacker 1983). Hacker shows, that the knowledge provision. For example double work can
planning phase of a task is the decisive phase, when be avoided and new knowledge is continuously created
existing knowledge is taken into account and used for (Probst & Buchel 1998).
the solution of a task. Thus, the more relevant knowl- In order to provide knowledge pro-actively, an IT
edge is available in the planning phase, the more poten- system has to act partly autonomous. The system
tial problems can be identi ed and potential mistakes has to know about the user's task and provide infor-
avoided. mation relevant for this task. Thus the system has to
There is also a reason why the system should act autonomously observe the user's actions, infer the user's
partly autonomous in providing information. In plan and search for information relevant to the user's
(Heinen & Dietel 1991) a model about information re- work context. This relevant information can then be
quired for the ful llment of a task is presented (Figure delivered to the user.
1). Heinen uses the term information, to which we will Another important aspect for knowledge acquisition
is the accordance of the personal context and the con- presented in (Schlichter, Koch, & Xu 1998). The char-
text of provided knowledge. The better knowledge is acteristics we focus on are:
personalized to the user's context the better it is ac-  Team: members know each other, have common goal,
knowledged and internalized by the user (Forgas 1985). involved in common processes
Information provided to the user should be person-
alized to the user's assumed quali cation and  Community: members do not know each other, mem-
interest. A negative e ect that can not be avoided, is bers have common interests but no common goal,
that if new knowledge contrasts an established belief of members are usually not involved in common pro-
the user, it is most probably not internalized. cesses
A human communication partner is undoubtedly the A user is member in a team and in one ore more commu-
most ecient and e ective knowledge source. One char- nities. Each of the knowledge source classes has their
acteristic that we would like to copy from a human com- own knowledge repository that is exclusively under con-
munication partner is graceful degradation when di- trol of the owners. We can see in Figure 2 that the
rect answers to a query are not possible. Human part- architecture is inherently distributed.
ners often still try to help and supply links to possible
solutions, which is much more agreeable than a response
like \0 hits found on your query".
Finally, the cognitive overload in using a KM IT sys-
tem has to be as little as possible. This means all cog-
nitive resources should be devoted to the acquisition of
knowledge and not to how to communicate with the IT
system. Thus we claim that high-level communica-
tion or maybe even anthropomorphism is required Communication

for a KM IT system. The user should be able to co-


operate with the system on a peer-to-peer basis and
communicate with the system in a discourse fashion
rather than giving explicit commands.
Information Management Architecture
We will present an IT infrastructure that incorporates Team member
the above derived functionalities. First, we motivate Team Community Agent
why we chose agent technology for the framework. We Community
will then give an overview over the structure of the
framework and afterwards explain the inner details of Figure 2: Relation of User, team and community
the framework in more detail.
Why agents ? The communication in and among communities will
A lot of the requirements above parallel what is pro- be explained later when we look in detail at the archi-
posed in (Foner 1993) as characteristics of a software tecture. First we would like to introduce our framework
agent system. Agents act partly autonomous and can in an overview (Figure 3).
thus pro-actively provide the user with required in- Information that is private to the user is adminis-
formation. The user gives strategic acting guidelines tered in the user agency, team relevant information is
to the agent instead of strict commands. This allows administered in the team agency and community rel-
on the one hand for a peer-to-peer-like cooperation be- evant information is administered in the community
tween the agent and the user and on the other hand for agency. The community agency is moreover the gate-
a graceful degradation in the performance of the task way from the team or the user to all communities in
by the agent. Agents are de ned to be personalizable, which a user is a member. Our architecture stays on
e.g. adaptive and adaptable to the user's needs. The a high level of abstraction. In the proposals for imple-
agent paradigm also provides for high-level interaction mentation it will be shown how existing KM IT support
between the user and the agents and is thus more suit- tools can be integrated in the framework.
able for knowledge exchange. Another advantage of the
agent paradigm is the loose coupling among functional The shared ontology is shared by all team members
components (agents). This assures an open architecture and provides a concept of the team knowledge domain.
and easy integration of future new components. The ontology on the one hand provides the basis for
consistent agent communication. On the other hand it
Framework structure is the scheme with regard to which existing knowledge
We have structured our framework according to three of users and knowledge required by users is uniquely
central knowledge source classes: user, team and com- identi ed. Moreover, it is the classi cation scheme for
munity. We use those terms according to what has been storage of information provided by the users and re-
User Agency knowledge of the user can be classi ed and respective
Profile Agent Knowledge Pump Agent Awareness Agent knowledge nodes can be marked as existent or non-
existent. The total set of these marks in the ontology
constitutes the user quali cation pro le.

document
knowledge
specified

nalized
perso-
request
The awareness agent analyzes the user's work con-
text. From observed user actions, a work context and
Service Agency a work plan are inferred. Both context and plan are
Information Broker Agent Directory Agent classi ed according to the ontology. This classi cation
is considered as the basis for the knowledge required by
the user to perform the current task.
document
Specified

Personalized document
request

nalized
Perso-

Specified request
The knowledge pump agent requests the work con-
text from the awareness agent. It further requests the
Team Agency user's knowledge pro le from the pro le agent. The dif-
Technology Agent ference of the two yields the knowledge still to acquire
Process Agent by the user to perform the current task. The knowl-
edge pump sends a request for these uniquely speci ed
Community Agency
knowledge nodes to the service agency.
Knowledge
Expertise Agent Community The service agency is the intermediate between the
Knowledge Firewall Agent Exchange
user agency and the other agencies. The service agency
consists of an information broker agent and a di-
rectory agent. The directory agent registers available
knowledge sources like the team agency or the commu-
Shared Ontology
nity agency. These knowledge sources do not necessar-
Knowledge Domain Model Process Domain Model
ily have to be existent in the exact fashion presented
here or may not exist at all. Or there may be more
Figure 3: Architecture of a KM IT system knowledge sources than the ones presented here. As
long as the sources are correctly registered with the di-
rectory agent, they are accessible to the information
trieval of this information. We will give a more detailed broker agent.
description of the usage of the ontology when the dif- The information broker agent receives the request for
ferent agencies are described. Since the shared team the speci ed knowledge nodes and looks up the avail-
ontology is not suitable for communication among all able knowledge sources with the directory agent. The
members of a community, the ontology shared by all information broker then requests the speci ed knowl-
members of a community is the set intersection of all edge nodes from the available sources. In our proposed
team ontologies. If the possibility of merging ontologies framework the sources are the technology agent and the
is provided, missing ontology parts can be acquired. process agent in the team agency as well as the exper-
Without going into implementation details, we con- tise agent in the community agency.
sider the ontology to conceptualize of a collection of
knowledge nodes that represent a uniquely identi ed The team agency includes two of the knowledge
\piece of knowledge". sources in our framework: the technology agent and
The shared ontology is divided into a knowledge the process agent. The technology agent administers
domain model and a process domain model. The explicit technical knowledge. The process agent admin-
knowledge domain model provides a concept of the rel- isters process experience. Each of these agents receives
evant explicit technical knowledge. The process domain a request for speci ed knowledge nodes in their domain.
model provides a concept of the tacit process knowledge The requested nodes may or may not exist. However, all
about processes common to a team. Since both parts found nodes are assembled into a document. The struc-
of the ontology are employed the same way, we will ture of this assembly is in the responsibility of the team
not distinguish between process and knowledge domain agents. The result of the assembly is a document that
model from now on. features the knowledge required by the user to perform
the current task, tailored to the user's quali cation.
The user agency consists of the pro le agent, the The community agency provides for the relational
knowledge pump agent and the awareness agent. network knowledge. The community agency consists
The pro le agent keeps a quali cation pro le of the of the expertise agent and the knowledge rewall
user. The pro le is based on the shared ontology. The agent.
shared ontology provides a concept of the team knowl-
edge domain. According to this concept, the existing The expertise agent is a knowledge source like the
technology agent and the process agent. The expertise user pro les can be designed and what possible appli-
agent administers knowledge about experts for certain cations are. We propose a user pro le that is based
knowledge sub-domains. The reference classi cation for on a stereotype. The stereotype is established in an
expertise is again the shared ontology. Since members initial user interview. From then on, the user pro le is
in a team according to our de nition know each other, dynamically extended through observation of the user's
expertise brokering is only needed in and among com- information request behavior. Empirical results showed
munities. It can be seen in gure 2 that the commu- that the time spent to read a document is correlated
nity agency is the gateway for communication among with the amount of the internalized knowledge from the
teams, be it within one or among di erent communities. document.
The expertise agent requests expertise knowledge from The functionality of a knowledge pump is shown
other community agencies. This knowledge exchange by the example of the XEROX Knowledge Pump in
can only take place if the respective ontologies have a (Borgho & Pareschi 1998).
non-empty intersection. This is a reasonable constraint, Another area for future research is the awareness
since one cannot expect an answer from someone who agent. Existing awareness infrastructures like NESSIE
does not know what one is talking about. (Prinz 1999) or Elvin (Fitzpatrick et al. 1999) sup-
The knowledge rewall agent controls what team- ply events from user observation. The missing link is
internal knowledge can be made available to non- plan recognition from those actions. We propose to
members of the team. Thus not all the team knowledge have the user himself semantically characterize di er-
is available to communities in which a team member ent workspaces. Upon switching of workspaces, a large
is also a member. An expertise agent may receive a pert of the work context is then already available to the
request for expertise knowledge, but the correspond- awareness agent. Another possibility is the integration
ing answer may be prohibited to leave the team by the of existing work ow systems, from which the currently
knowledge rewall agent. The knowledge rewall agent worked process could be extracted.
also has the task to request technical and process knowl- The service agency contains standard functionality
edge from other teams on request of the information in distributed systems.
broker agent in the service agency.
The team agency and the community agency
Existing Implementation Concepts and The main issue here is storage and reuse of informa-
tion modules. A promising approach seems to be usage
future research of XML for information representation (Bray, Paoli,
We will brie y introduce some possible implementation & Sperberg-McQueen 1998), (null W3C 1999). The
concepts for our framework. With this we also identify exible syntax approach and the possibility of seman-
areas for future research. We will give some examples tic meta-tagging makes XML predestined for informa-
for how existing KM IT tolls can be integrated in our tion reuse and dynamic assembly. In (Su, Freitag, &
framework. Brossler 1999), an XML-based teachware system is pre-
Infrastructures for software agents have been exten- sented. Information modules there can be reused, ex-
sively researched and there are a number of existing ibly and dynamically structured in new documents and
solutions. A comprehensive collection of articles on the presentation style of the document can be adapted.
multiagent systems can be found in (Weiss 1999). For Existing document management systems for example,
agent communication a number of Agent Communica- could be integrated as means of storage for information
tion Languages (ACL) have been proposed. KQML in the team agency and the community agency.
(Finin et al. 1992) is targeted towards knowledge ex-
change. The FIPA ACL (Steiner 1998) is extensible
for di erent purposes and well-maintained due to the
Conclusion
powerful industrial background of FIPA. We found that existing KM e orts either su er from
The shared ontology The core of our architecture patchwork IT solutions or from non-acceptance of IT
is the shared ontology. The Ontolingua system (Fikes, solutions that were not designed with a holistic per-
Farquhar, & Rice 1997) is a powerful tools to design spective on KM. We derived core functionalities of KM
ontologies. In (Bullock & Goble 1998) a semantic net is IT support by taking a holistic perspective on KM
built on the basis of a description logic to conceptualize and related research. We identi ed the need for high-
objects in the domain of tourism. This is an approach level communication to reduce the cognitive usage over-
that could be easily transferred for an ontology imple- load. Moreover, information needs to be provided au-
mentation. In (Gruber 1993), a number of design prin- tonomously and pro-actively. And nally the supplied
ciples for ontologies are introduced. We consider the information must be tailored to the requirements of
shared ontology and the user agency the parts of our the user, especially his quali cation. Those compo-
architecture, where most research still has to be done. nents were transferred into an IT system framework.
Agent technology was identi ed as the key to many of
The user agency Kobsa presents in (Kobsa & the identi ed needs. We also identi ed user, team and
Wahlster 1989) and in (Fink, Kobsa, & Nill 1996) how community as the important concepts for the scope of
knowledge administration. We structured our frame- Foner, L. N. 1993. Whats an agent anyway? a socio-
work according to these scopes as well as according to logical case study. Technical report, MIT Media Lab.
derived classes of knowledge to administer (technical Forgas, J. P. 1985. Interpersonal behaviour. Pergamon
explicit knowledge, tacit process experience and rela- Press.
tional network knowledge). Gruber, T. R. 1993. Toward principles for the design
We brie y discussed existing approaches for possi- of ontologies used for knowledge sharing. Technical
ble implementation of our framework and showed how report, Stanford University.
existing tools can be integrated in our framework. Pro- Hacker, W., ed. 1983. Kognitive und motivationale
posals for implementation of user pro ling, document Aspekte der Handlung, volume 22. Huber.
personalization and awareness services were given. We
further identi ed areas for future research within the Heinen, E., and Dietel, B., eds. 1991. Industriebetrieb-
scope of our framework. slehre, volume 9. Gabler.
This paper sets a basis for future research in IT- Kobsa, A., and Wahlster, W., eds. 1989. User Models
related issues of KM. The importance of a holistic per- in Dialog Systems. Springer Verlag, London.
spective for requirement identi cation was shown. The Nonaka, I., and Takeuchi, H. 1995. The Knowledge-
framework we proposed is open for future extensions. Creating Company. Oxford University Press.
Acknowledgments. This research is sponsored by North, K.; Probst, G.; and Romhardt, K. 1998. Wis-
the German Governmental Research Association sen messen - ansatze, erfahrungen und kritische fragen.
(DFG). We would like to thank Jurgen Koch as well ZfO - Zeitschrift Fuhrung und Organisation 3/1998:p.
as all members of the Arbeitskreis Wissensmanagement 158{166.
(KM Workgroup) for many fruitful discussions. A spe- null W3C. 1999. Xml faq. Webpage.
cial thank goes to Christine Erlach for many useful Polanyi, M. 1958. Personal Knowledge. Chicago.
hints and the contribution of psychological background
knowledge. Prinz, W. 1999. Nessie: An awareness environment
for cooperative settings. In ECSCW99.
References Probst, G. J., and Buchel, B. S. 1998. Organisationales
Lernen: Wettbewerbsvorteil der Zukunft. Gabler.
Borgho , U. M., and Pareschi, R., eds. 1998. Informa- Su, C.; Freitag, B.; and Brossler, P. 1999. Meta-
tion Technology for Knowledge Management. Springer modeling for web-based teachware. In International
Verlag, Berlin. WWWCM 99. Springer LNCS.
Bray, T.; Paoli, J.; and Sperberg-McQueen, C. 1998. Schlichter, J.; Koch, M.; and Xu, C. 1998. Awareness
Extensible markup language (xml) 1.0. Webpage. - the common link between groupware and commu-
Bullinger, H.-J.; Worner, K.; and Prieto, J. 1998. nity support systems. In Ishida, T., ed., Community
Wissensmanagement - Modelle und Strategien fur die Computing and Support Systems, pp. 77{93. Springer
Praxis. Springer. p. 21{39. Verlag.
Bullock, J., and Goble, C. 1998. Tourist: The applica- Singley, M., and Anderson, J. 1989. The transfer of
tion of a description logic based semantic hypermedia cognitive skills. Cambridge, Massachusetts.
system for tourism. In ACM Hypertext 98, p. 132{141. Steiner, D. 1998. Die pa-initiative fur agenten-
ACM. standardisierung. it und ti - Informationstechnik und
Fikes, R.; Farquhar, A.; and Rice, J. 1997. Tools for Technische Informatik 40:pp. 46{49.
assembling modular ontologies in ontolingua. Tech- Sveiby, K. E. 1998. The new organizational wealth.
nical report, Stanford University Knowledge Systems Berrett-Koehler Publications.
Laboratory. Weiss, G., ed. 1999. Multiagent Systems - A Modern
Finin, T.; Weber, J.; Wiederhold, G.; Genesereth, M.; Approach to Distributed Arti cical Intelligence. MIT
Fritzson, R.; McKay, D.; McGuire, J.; Pelavin, P.; Press.
Shapiro, S.; and Beck, C. 1992. Speci cation of teh
kqml agent-communication language. Technical re-
port, Enterprise Integration Technologies, Palo Alto,
CA.
Fink, J.; Kobsa, A.; and Nill, A. 1996. User-oriented
adaptivity and adaptability in the avanti project. In
Designing for the Web: Empirical Studies.
Fitzpatrick, G.; Mans eld, T.; Kaplan, S.; Arnold, D.;
Phelps, T.; and Segal, B. 1999. Instrumenting and
augmenting the workaday world with a generic noti -
cation service called elvin. In ECSCW99.

Das könnte Ihnen auch gefallen