TOWARDS A SELF-ADAPTIVE MULTI-AGENT APPROACH
FOR ENHANCING THE QUALITY OF SERVICE PROVIDED BY
OPEN INFORMATION SYSTEMS
Valérie Camps and Pierre Glize
IRIT- Université Paul Sabatier – 118 route de Narbonne 31062 Toulouse Cedex 4, France
Keywords: Web site and services personalisation, searching and browsing, emergent quality of services, adaptive multi-
agent systems, self-organisation.
Abstract: Current information systems are plunged into highly dynamical environments which produce occurrences of
unpredictable situations. This dynamics combined with the inherent geographical and functional distribution
of such systems, make usual adaptation techniques which are global or dependent of the intended global
function realised by the system, unsuitable. Our contribution concerns a partial instantiation of a local
adaptation method, based on adaptive multi-agent systems, to manage the QoS of information systems. This
management is done according to two points of view addressed in an integrate way: a quantitative one and a
qualitative one. First obtained results, showing the benefits of cooperation to the adaptation of such systems,
are then discussed.
1 INTRODUCTION
We consider open systems (Hewitt, 1982) as
evolving systems composed of dynamic entities and
plunged into a dynamic environment. In these
conditions, how designers of open information
systems (IS) can guarantee to humans who use them
a given quality of service (QoS)? The usual response
is to propose infrastructure, middleware, norms,
protocols, Web Services. Unfortunately, this is a
never-ending process in open
systems: heterogeneity, incompleteness and
unforeseeable situations are inescapable and have to
be taken into account. We propose a quite new
different approach by first agentifying all the
components of an open IS and then giving them the
capability to collectively converge dynamically and
in real-time towards the optimal QoS. This original
approach optimises the QoS of an IS by only taking
into account individual characteristics of its
components, instead of improving or finding new
criteria to deal with its QoS as well as creating new
tools to manage it. Generally speaking QoS has two
constituents: (i) qualitative (functional) properties,
defining how well the retrieved information matches
the intended information such as precision, recall
and noise and (ii) quantitative (non functional)
properties, ensuring an effective flow in terms of
end-to-end delay and including properties such as
security, breakdown, interoperability, bandwidth.
This paper presents how the agentification
process is able to tackle the above constituents of
QoS in open IS from two experiments: a quantitative
one, dealing with end-users and services mapping
according to a need, and a qualitative one, dealing
with a learning algorithm prefiguring the
construction of end-users and services profiles. This
work is based on the AMAS (Adaptive Multi-Agent
Systems) approach allowing the design of complex
systems that can be incompletely specified and for
which an a priori known algorithmic solution does
not exist. First innovative aspects of our approach
are then presented. We conclude by proposing a
guide to study properties of QoS in open IS tackled
with an emergent problem solving approach.
295
Camps V. and Glize P. (2007).
TOWARDS A SELF-ADAPTIVE MULTI-AGENT APPROACH FOR ENHANCING THE QUALITY OF SERVICE PROVIDED BY OPEN INFORMATION
SYSTEMS.
In Proceedings of the Third International Conference on Web Information Systems and Technologies - Web Interfaces and Applications, pages 295-301
DOI: 10.5220/0001286802950301
Copyright
c
 SciTePress

2 QUANTITATIVE QoS
MANAGEMENT
Adaptive profiling of end-users/services is an
inescapable approach for an IS dealing with dynamic
user-service mapping, but is not sufficient. An
adaptive response to problems induced by the
dynamics and the heterogeneity of such systems
(such as workload, failures, interoperability of the
components, as well as the integration of new
services...) becomes also necessary. For that, we
propose a full agentification of IS components. The
functional architecture we propose to tackle this
problem and a cooperative protocol between the two
upper agents levels are given in this paragraph. The
objective of this protocol is to put in touch an entity
having a task to achieve (expressed by a request)
with entities able to answer (a relevant service).
2.1 Functional Architecture
This part uses two types of agents, which
respectively belong to a different level of the
architecture of our system (figure 1).
The first type of agent is called representative
agent. A representative agent acts on behalf of the
end-user or the service it represents in order to solve
a submitted request. A user agent has to seek for
service agents that fit as well as possible the needs
expressed by its end-user (calculus for Grid
Computing, service for Web Services…).
Conversely, a service agent, during a publicity
campaign for example, can have to find user agents
likely to be interested by the service it proposes.
The second type of agent is called site agent. A
site agent helps each representative agent it
contains, locally or remotely (by communicating
with other site agents to find new relevant
representative agents). This agentification is
required because end-users and services are
numerous and geographically distributed. As it is
unrealistic to gather them into a single site, we
consider a distributed IS as composed of several site
agents which respectively contain numerous
representative agents.
Representative agents and site agents follow the
same “cooperative” protocol (this concept is
explained in §4). They interact according to
representations they have on other agents’ skills.
This protocol can be summarized by the five
following steps. Ideally, when (i) the received
message is totally and without ambiguity understood
by the agent, it processes it. When (ii) an agent
cannot associate a meaning to the received message,
it sends the message towards an agent it considers
relevant for the resolution (this action is called
“restricted relaxation”). Thanks to this action, the
original sender agent can have the opportunity to
become acquainted with a new agent. When (iii)
only a part of the received message has a meaning
for the agent, it returns to the sender a partial answer
corresponding to the understood part and it sends the
remainder to an agent it considers qualified
(restricted relaxation). When (iv) the received
message has several meanings for the agent, it
returns the message to the sender for clarification.
When (v) two agents want to reach a third one
proposing a limited resource and their requests
exceed the offer, they are faced with a situation of
A
A
A
A A
A
A
A
A
Representative Agent
Site a
g
en
t
Service
Real e
n
tity/service
A
A
A
A
A
A
A
A
Document Agent
A
Term Agent
Descriptor-Term Agent
Specialization/Generalization link
Figure 1: Functional architecture of the proposed system.
Service
Se
r
v
i
ce
Service
WEBIST 2007 - International Conference on Web Information Systems and Technologies
296

conflict. In this case, the third agent guides one of
the former agents towards another having similar
resource. All the agents are encapsulated by the
same behaviour. An agent can thus recommend
agents having similar competences when it is
overloaded. This behaviour is relevant even if
involved agents propose concurrent sale services
because the global QoS increases and all services
can potentially take benefit of it.
According to this protocol, each agent aims at
promoting interactions with agents having similar
centres of interest. Conversely, it tries to weaken,
even to remove, links with agents having different
centres of interest. Furthermore, this part has also to
take it into account and self-adapts consequently to
the centres of interests of the real represented actors.
2.2 Cooperation for End-Users and
Services Connection
The contribution of cooperation for end-users and
services mapping was already highlighted in several
applications (Gleizes, 2002), (Link-Pezet, 2000). But
these applications only focused on qualitative
considerations: skills of other involved agents.
Obtained results at the end of these projects pointed
out the necessity of better taking into account the
needs of end-users and proposed services. As
obtained results in these applications were closely
related to the chosen mode of representation of the
centres of interest of involved actors, we decided to
check the contribution of the cooperation in a more
general context, by also taking into account
quantitative properties.
We then made a simulation of a network of
heterogeneous, distributed and dynamic ISs (Grid
Computing, Web Services and Peer-to-Peer),
implementing temporal resources, processes and
requests to be solved (Cabanis, 2006). The
cooperative protocol previously presented was
instantiated to this context by taking into account not
only the supposed skills of involved agents but also
several criteria (such as CPU performance, storage
capacities, standards and bandwidth). In this
simulation, representations of agents are
implemented by using measurements of need
(standards and access rights in Web Services), of
probability (for the reliability of the services) and of
weighted averages for apparent performances. This
simulation, developed in JavAct (see
http://www.javact.org for more details), consists in
100 agents, 80% of which are devoted to Grid
Computing (GC) calculus. Initially, the IS is
represented by a graph of agents randomly
connected. This graph evolves according to
interactions between agents. 90 requests of GC
calculus are submitted each second to different
agents of the system. According to the previously
presented protocol, each task/request can be relaxed
a limited number of times (4 times in this
simulation). Beyond this number, the task/request is
removed and the sender representative agent
considers its request as being without response after
a given time limit (Time-out). It then adjusts
consequently its representations on the
representative agent to which it sent the request.
Obtained results (see figure2) show a progressively
decreasing number of relaxations and a decreasing
number of Time-out (unsolved requests/tasks)
during the system functioning. These results mean
that gradually each agent finds its right place in the
organisation in spite of unforeseeable events that can
occur during the system functioning. In the second
curve an asymptotic limit to 20% of time-out can be
seen. It is reached when all agents devoted to GC are
busy; so the system tends towards its optimality.
According to the QoS, these preliminary results
show well that the network, as a collective, adapts
itself to the characteristics of each entity, only by
local perception of criteria and treatments which are
independent of any global cost function knowledge.
3 QUALITATIVE QoS
MANAGEMENT
A representative agent is composed of two types of
representations (also called beliefs or profile):
Time-out proportion evolution
0%
20%
40%
60%
80%
1 101 201 301 401 501 601 701
Relaxations number by request
0
1
2
3
4
5
1 101 201 301 401 501 601 701
Relaxation number
Logarithmic approximation
Figure 2: Cooperation contribution on QoS
Time (sec)
Time (sec)
Time-out proportion evolution
0%
20%
40%
60%
80%
1 101 201 301 401 501 601 701
Relaxations number by request
0
1
2
3
4
5
1 101 201 301 401 501 601 701
Relaxation number
Logarithmic approximation
Figure 2: Cooperation contribution on QoS
Time (sec)
Time (sec)
Figure 2: Cooperation contribution on QoS.
TOWARDS A SELF-ADAPTIVE MULTI-AGENT APPROACH FOR ENHANCING THE QUALITY OF SERVICE
PROVIDED BY OPEN INFORMATION SYSTEMS
297

representations about agents already contacted
during previous researches and representations about
the end-user or the real service it represents. A
representative agent is supposed described by a set
of textual data (documents such as HTML pages for
example). After the lemmatization of this set of
documents, the objective consists in extracting from
these documents a set of descriptors, i.e. a signature
characterising as well as possible their "semantic"
content, and therefore, the centres of interest
(profile) of the represented end-user or service.
3.1 Functional Architecture
The qualitative QoS part is mainly composed of two
types of agents, which respectively belong to a
different level of the architecture (figure 1).
The first type of agent is called term agent. A
term agent represents the lemmatized version of a
word initially contained in a particular document.
Each term agent possesses a confidence degree
computed in real-time according to the evolution of
the execution context and the dynamics of the
environment. The more representative a term of the
document is, the higher its confidence degree is.
Beyond a given threshold, a term agent becomes a
descriptor-term agent. The objective of a descriptor-
term agent is to be connected/not connected to other
descriptor-term agents semantically or contextually
close/distant. Thus, a descriptor-term agent takes
part in the construction of a terminological network
which will be used to know the centres of interest of
a representative agent. To do that, we consider each
term agent obtained at the end of the lemmatization
of the documents characterising a representative
agent. If it is located in the neighbourhood (two term
agents are neighbours if they are near one another in
a document) of a great number of distinct term
agents, it does not take part in the highlighting of
descriptor-term. Its confidence is then reduced.
The second type of agent is called document
agent. A document agent represents a particular
document describing an end-user or a service. A
document agent is initially connected to all term
agents and descriptor-term agents composing it. It
will then only keep links with descriptor-term agents
which characterise it and which make up its
signature. The objective of a document agent is to
allow the highlight of semantic or contextual
features which describe the centres of interest of the
represented entity, while taking into account their
evolution. To do that, we consider that when two
document agents are similar, they deal with close
problematic. In that case, descriptor-term agents
common to these two documents and having a low
confidence degree must change place in the
organisation and try to connect themselves to
descriptor-term agents taking part in the signature of
the actual representative agent. The treatment is
symmetrical when considered documents are
dissimilar. The confrontation of documents
(similarity/dissimilarity) can be realised at various
levels of the mapping process: when a modification
of the real entity is made, when a task or a request is
submitted to the IS, when an end-user gives a
feedback on the quality of the connection relation
(QoS) or when an end-user explores a document
(during an information retrieval).
Two types of links exist between descriptor-term
agents. The first one is called contextual closeness
link; it connects two descriptor-term agents having
similar contextual interests. It is directed from an
agent A towards an agent B where A is contextually
supposed to be more specific than B. The second
one is called contextual identity link: it connects two
descriptor-term agents having similar contextual
interests with a bidirectional contextual closeness
link. This mean that involved descriptor-term agents
are considered as similar in the current context.
3.2 Cooperative Profiling
We evaluated the feasibility and the relevance of our
adaptive, local, and independent of semantic
treatment (except the lemmatization) algorithm, on
the design of profiles (Czerny, 2006). We made first
experiments, based on a corpus of documents
(around fifty) resulting from RFIEC platform
(http://www.irit.fr/RFIEC). This corpus was
composed of articles of the daily French newspaper
"Le Monde" of the year 1994 (dealing with the
architecture in Berlin, the drug in Holland and the
French conscientious objectors) as well as a list of
correspondences "Requests - Documents".
A subset of the terminological network we
obtained is showed in figure 3. It was exclusively
built according to the local behaviours previously
presented (for a better understanding descriptor-
terms had been translated from French into English).
It presents interesting characteristics, notably the
absence of meaningless terms, the existence of links
between semantically/contextually close descriptor-
term agents and a kind of semantic proximity in the
neighbourhood of some descriptor-term agents (for
example around the descriptor-term "narcotic"). In
this example, various links between the descriptor-
term agent "netherlands" and descriptor-term agents
such as "narcotic", "cannabis", "drug", "methadone"
WEBIST 2007 - International Conference on Web Information Systems and Technologies
298

and "drug addict" can be distinguished. Interests of
the active representative agent on the Netherlands
are then supposed to be related to drugs. If some
links are relevant, others (such as the descriptor-
term agent "according to" in the network and
associated links) are less pertinent. A
complementary work remains to be done on
relations connecting two descriptor-term agents.
4 DISCUSSION
For the last few years, user profiling has been
becoming a research field of topical interest. This
craze has its origins in information retrieval field.
Nowadays, the number of responses provided by
search engines to a user’s request remains high;
locating relevant information in the list of returned
documents is not an easy task and needs a
considerable amount of time. User modelling can
contribute to several steps of the research process
notably for exploring information sources and
delivering only the most relevant documents to a
user. (Daniels, 1986) contrasts two classes of user
model: quantitative and empirical models which
study the external behaviour of a user by observing
his interactions with the system, and analytical and
cognitive models which are interested in modelling
the internal behaviour of a user and try to identify
the knowledge and the cognitive processes used. Our
work relates to the second point.
4.1 Qualitative QoS
User profile acquisition can be performed in an
explicit way, by collecting the information provided
by a user via the system interface (selection of
topics, definition of attributes, explicit judgment on
the relevance of document...), or in an implicit and
dynamic way, by observing his behaviour when he is
interacting (bookmark saving, link selection, total
time spent on a page...) with the system (Lieberman,
1995), (Albayrak, 2005). Most of existing systems
use vectorial representations coupled with standard
weighting schemes to draw up user profiles.
Semantic representations are sometimes used too.
They display relations between the units of
information characterising the profile by proposing a
hierarchy of concepts. They are generally based on
ontologies (Baziz, 2005) which confer a quite
relative adaptation because they are dependent of a
given domain. These approaches are sometimes
coupled with techniques that take into account the
evolution of the profile. Some systems employ
learning algorithm adopted from neural networks or
genetic algorithms (Menczer, 1997). Most of these
approaches, except (Moukas, 1997), do not address
their effectiveness to adapting to changing user’s
interests. More recent works try to take into account
a temporal dimension (short/average/long terms
interests) (Kilfoil, 2005) or information related to
the context of the user (Bottraud, 2004). But these
adaptive approaches rest on global solutions or base
their reasoning on the expected result of the system,
which makes them not easily applicable for
simultaneously managing multiple criteria and
unforeseen situations apparition.
4.2 Quantitative QoS
Researches addressing the quantitive QoS problem
of applications deployed in large-scale distributed
and heterogeneous environments are quite new. As
(Kalogeraki, 2005) says “The inherent ad-hoc nature
of these systems makes it difficult to meet the Quality
of Service (QoS) requirements of the distributed
applications, thus having a direct impact on their
scalability, efficiency and performance”. For
example in a Peer-to-Peer network (Drougas, 2006)
proposes an adaptation mechanism, which trades off
service quality level with resource usage. (Cuenca-
Acuena, 2004) proposes cooperative agents to gather
information about the system state and services.
Adaptive QoS is also required in a heterogeneous
(wired and wireless) environment. In this context
(Chowdhury, 2002) proposes a collaborative
framework for adaptive QoS management to support
interactive information sharing among distributed
and heterogeneous clients, while (Angin, 1998)
chooses a mobile middleware toolkit to adapt mobile
Figure 3: A subset of the obtained terminological
network
TOWARDS A SELF-ADAPTIVE MULTI-AGENT APPROACH FOR ENHANCING THE QUALITY OF SERVICE
PROVIDED BY OPEN INFORMATION SYSTEMS
299

services. A survey for an adaptive QoS based on
middleware solutions is given in (Duran-Limon,
2004). (Chen, 2004) gives also a quite complete
survey of approaches for improving the QoS in
wireless sensors networks.
An important point that comes out of these works
is the need of adaptive mechanisms to tackle the
QoS of open and distributed systems.
4.3 Towards an Emergent Optimal
QoS
We believe as well, that faced with the diversity of
criteria to take into account (CPU performance,
capacities storage, standards, bandwidth…), the
required QoS cannot be checked and managed by an
external supervision. The entities of the system
should be autonomous and adapt themselves locally
to environmental changes, according to what they
perceive and their internal state. Thus, the learning
phase is a never-ending process.
We define a system as being functionally
adequate if it produces the function for which it was
conceived, according to the viewpoint of an external
observer knowing its finality. We consider the
functional adequacy problem of an open IS as a QoS
optimisation one. To reach this functional adequacy,
it had been proven (Camps, 1998) that each
autonomous agent composing an AMAS and
following a cycle composed of three steps
(perception/decision/action) must keep relations as
cooperative as possible with its social (other agents)
or physical environment. The definition of
cooperation we use is not a conventional one (simple
sharing of resources or common work). Our
definition is based on three local meta-rules the
designer has to instantiate according to the problem
to be solved: (c
per
) every signal perceived by an
agent must be understood without ambiguity, (c
dec
)
information coming from its perceptions has to be
useful to its reasoning, (c
act
) this reasoning must lead
the agent to make actions which have to be useful
for other agents and the environment. If one of this
meta-rule is not checked, the agent is faced to a
"Non Cooperative Situations" (NCS). A NCS can be
assimilated to an "exception" in traditional
programming. Our approach is a proscriptive one
because each agent has first of all, to anticipate, to
avoid and to repair a NCS. A NCS occurs when at
least one of the three previous meta-rules is not
locally verified by an agent. Different generic NCSs
can then be highlighted: incomprehension and
ambiguity if c
per
is not checked, incompetence and
unproductiveness if c
dec
is not obeyed and finally
uselessness, competition and conflict when c
act
is not
checked. This approach has great methodological
implications: designing an AMAS consists in
defining and assigning cooperation rules to agents.
In particular, the designer, according to the current
problem to solve, has (i) to define the nominal
behaviour of an agent then (ii) to deduce the NCSs
the agent can be confronted with and (iii) finally to
define the processing the agent has to perform to
come back to a cooperative state.
This approach is the basis of the QoS
management we propose. In the quantitative QoS
management, a protocol composed of five steps had
been defined. The first step is the nominal behaviour
(the agent is in a cooperative state). In the fourth
other steps, the agent is faced to an NCS. More
precisely, in the second step the agent is faced to a
total incomprehension, in the third one, it is faced to
a partial incomprehension, in the fourth one to an
ambiguity and in the fifth one to a conflict. In the
same way, in the qualitative QoS management, two
NCSs had been defined: (i) a uselessness NCS when
a term agent is in the immediate neighbourhood of a
great number of distinct term agents in a document
and (ii) a unproductiveness NCS when two
descriptor-term agents common to two semantically
close documents have a weak confidence degree. In
these two cases, involved agents do not take part in
the construction of the terminological network.
Behaviours agents have to carry out when they
are faced to an NCS are given for each underlined
NCS. These behaviours lead to a local
reorganisation of interaction links between involved
agents. These agents do not have a view of the
global system and do not base their reasoning on the
expected collective function realised by the system.
5 CONCLUSION AND
PERSPECTIVES
Because of their complexity, current ISs require new
approaches to apprehend volatility, dynamics and
opening problems. Traditional adaptive approaches,
based on the expected function of the system are not
easily applicable (even unsuited) to take into account
the unforeseeable environmental constraints. Our
contribution is a definition of a local adaptive
approach, based on permanent cooperative
interactions between entities composing the system.
We presented its partial instantiation to
quantitative and qualitative QoS managements. First
obtained results as well as the contribution of the
cooperation on the system adaptation had been
displayed. These encouraging results convinced us
WEBIST 2007 - International Conference on Web Information Systems and Technologies
300

to study thoroughly the use of the AMAS for the
QoS in open IS. Several tasks still remain to be
realised: (i) to improve the learning algorithm which
associates a signature to a set of documents
describing a real entity and to extend its use with the
simultaneous representation of several profiles (an
agent must have an image of already contacted
agents); (ii) to implement the interrogation of the
profile built and then to integrate this learning
algorithm into the general process to determine
relationships; (iii) to study the use of the built
terminological network and its relations to allow the
expansion of requests to disambiguate a request
submitted by an end-user to the system; (iv) to
agentify real services/users with cooperative
behaviours to obtain truly generic and adaptive
networks.
All the researchers in IS consider implicitly or
explicitly, that improving the QoS is a multi-
criterion and dynamic optimisation problem. It is
also our case, but we consider, moreover, that
theoretical limitations of the usual algorithms of
optimisation lead ineluctably to a reduction of this
QoS progressively with the increasing complexity of
such systems. New ways based on emergent
problems solving can reverse this tendency.
REFERENCES
Albayrak S., Wollny S., Varone N., Lommatzsch A.,
Milosevic D., 2005. Agent Technology for
Personalized Information Filtering: The PIA-System,
in the 20th ACM Symposium on Applied Computing,
Agents, Interactions, Mobility, and Systems (AIMS)
track, Santa Fe, New Mexico, USA, pp. 54-59.
Angin O., Campbell A. T., Kounavis M. E., 1998. The
Mobiware Toolkit: Programmable Support for
Adaptive Mobile Networking, IEEE Personal
Communications Magazine.
Baziz M., Boughanem M., Aussenac-Gilles N., 2005.
Semantic Networks for a Conceptual Indexing of
Documents in IR, in the 7th International Symposium
on Programming and Systems (ISPS), Alger.
Bottraud J-C, 2004. Un assistant adaptatif pour la
recherche d'information : AIRA (Adaptative
Information Retrieval Assistant), Thèse de l'Université
Joseph Fourier.
Cabanis V., 2006. Etude de la dynamique auto-
organisationnelle du Web fondée sur l'activité
coopérative de ses composants, Master of research
report of Paul Sabatier University.
Camps V., Gleizes M.P., Glize P., 1998. A self-
organization process based on cooperation theory for
adaptive artificial system, in 1st Int. Conference on
Philosophy and Computer Science “Processes of
evolution in real and Virtual Systems”, Krakow.
Chen D., Varshney P. K., 2004. QoS Support in Wireless
Sensor Networks: A Survey, Proc. of the 2004
International Conference on Wireless Networks
(ICWN 2004), Las Vegas, Nevada, USA, June 21-24.
Chowdhury R., Bhandarkar P., Parashar M., 2002.
Adaptive QoS Management for Collaboration in
Heterogeneous Environments, in Proc. of the 16th Int.
Parallel and Distributed Processing Symposium
Cuenca-Acuna F., Nguyen T., 2004. Self-managing
federated services, in Proc. of the 23rd IEEE Int.
Symposium on Reliable Distributed Systems (SRDS
2004).
Czerny J., 2006. Apprentissage de Représentations
Locales dans un Système Multi-Agent de Recherche
d'Information, Master of research report of Paul
Sabatier University.
Daniels J.P., 1986. Cognitive models in information
retrieval, an evaluation review, Journal of
documentation, 42(4), 272-304.
Drougas Y., Repantis T., Kalogeraki V., 2006. Load
Balancing Techniques for Distributed Stream
Processing Applications in Overlay Environments, in
9
th
IEEE Int. Symposium on Object and Component-
Oriented Real-Time Distributed Computing
(ISORC'06).
Duran-Limon H.A., Blair G.S., Coulson G., 2004.
Adaptive Resource Management in Middleware: A
Survey, in IEEE Distributed System Online - IEEE
Computer Society, Vol. 5, No. 7
Gleizes M-P, Glize P., 2002. ABROSE: Multi Agent
Systems for Adaptive Brokerage, in 4th International
Bi-Conference Workshop on Agent-Oriented
Information Systems (AOIS), Toronto, Ontario.
Hewitt C., De Jong P., 1982. Open Systems, AIM-691,
ftp://publications.ai.mit.edu/ai-publications/pdf/AIM-
691.pdf
Kalogeraki V., Chen F, Repantis T, Zeinalipour-Yazti D,
2005. Towards Self-Managing QoS-Enabled Peer-to-
Peer Systems, LNCS Volume 3460
Kilfoil M., Ghorbani A.A., 2005. SWAMI: Searching the
Web Using Agents with Mobility and Intelligence, in
Canadian conference on AI, pp 91-102.
Lieberman H., 1995. Letizia: An Agent That Assists Web
browsing in Proceedings of the International Joint
Conference on Artificial Intelligence, Canada.
Link-Pezet J., Gleizes M-P, Glize P., 2000. FORSIC: a
Self-Organizing Training System, in International
ICSC Symposium on Multi-Agents and Mobile
Agents in Virtual Organizations and E-Commerce,
MAMA'2000, Wollongong, Australia.
Menczer F., 1997. ARACHNID: Adaptive Retrieval Agents
Choosing Heuristic Neighborhoods for Information
Discovery, Proc. of the Fourteenth Int. Conference on
Machine Learning.
Moukas A., 1997. User Modeling in a MultiAgent
Evolving System, in Proc. of Workshop on Machine
Learning for User Modeling, 6
th
Int. Conference on
User Modeling, Chia Laguna, Sardinia.
TOWARDS A SELF-ADAPTIVE MULTI-AGENT APPROACH FOR ENHANCING THE QUALITY OF SERVICE
PROVIDED BY OPEN INFORMATION SYSTEMS
301