PERSONALIZATION OF A TRUST NETWORK
Laurent Lacomme, Yves Demazeau
Laboratoire d’Informatique de Grenoble, Grenoble, France
Valérie Camps
Institut de Recherche en Informatique de Toulouse, Toulouse, France
Keywords: Trust, Agent network, Personalization, Social network.
Abstract: Trust and personalization are two important notions in social network that have been intensively developed
in multi-agent systems during the last years. But there is few works about integrating these notions in the
same network of agents. In this paper, we present a way to integrate trust and personalization in an agent
network by adding a new dimension to the calculus of trust in the model of Falcone and Castelfranchi,
which we will call a similarity degree. We first present the fundamental notions and models we use, then the
model of integration we developed and finally the experiments we made to validate our model.
1 INTRODUCTION
From Web Services to experimental negotiating
agendas, many multi-agent systems have been
developed to implement links between people or
organizations in order to enable them to interact
indirectly through agents that represent them. In
such social networks, each agent stands for a person
or a group of people.
These networks have often some particular
properties. The first we are interested in is openness.
In open networks, agents can be added or removed
from the network at any time. This implies that the
network evolves, while each agent needs to adapt its
own behavior to the appearance or disappearance of
partners. The second property is partial
representation. In social networks, agents often only
have little knowledge about others and about the
network itself. For instance, when an agent is added
to the network, it usually only knows a few other
agents that we call its neighborhood. A third
interesting property is heterogeneity. That is, in such
networks, agents are not always homogeneous.
Every agent can have individual skills that others do
not have, and each agent is free to cooperate or not
with known agents. So each agent has to choose
cleverly its partners in this kind of networks,
because these partners must fulfill some
requirements for the partnership to be useful.
Hence such networks need some protocols for the
agents to be able to act correctly while knowing only
a few facts about a constantly evolving environment.
One way to fulfill this requirement is to add a trust
model to their reasoning abilities. This trust model
enables them to take decisions, such as which agent
to ask for doing a task, from the little knowledge
they have. This is done by first computing probable
behaviors of others and the results of such behaviors,
and then selecting the best ones for the agent.
On another hand, as the agents in these networks
are used to represent human beings, users often want
to have some control over them. Indeed, when
agents are faced to choices, their reactions should be
the closest possible to the users’ own preference.
One way to realize it is to add to the agents’
reasoning methods a personalization model, which
checks for alternatives and selects the one the user
would prefer.
As a result of the two previous remarks, there is a
need to include both trust and personalization
models. But as both are reasoning methods that can
lead to contradictory conclusions, we need a way to
integrate them in the agent’s global reasoning
protocol. From our best knowledge, such an
integration does not seem to exist yet.
In the remaining of this paper, we will first
present the notions of trust and personalization in a
network of agents, the theoretical criteria we will
408
Lacomme L., Demazeau Y. and Camps V. (2009).
PERSONALIZATION OF A TRUST NETWORK.
In Proceedings of the International Conference on Agents and Artificial Intelligence, pages 408-415
DOI: 10.5220/0001659204080415
Copyright
c
 SciTePress

develop on these notions for our integration work
and the trust model we chose as a foundation to our
integration model. Then, we will describe the
integration criteria and the solution we are
proposing. And finally, we will present an
experimental validation of our solution.
2 POSITIONING
There are many ways for agents to represent and
compute trust they have in other ones, and there are
also many ways for them to represent and exploit
user’s preferences. So, to understand how trust and
personalization should interact in an agent’s
reasoning schema, we firstly need to describe what
they are, how they work, and what their different
possible models are.
2.1 Trust in Agent Networks
A trust model describes how an agent can use its
past experience and others’ experience to take
decisions about future plans. It involves a facts
storing and a reasoning method over this memory.
As we are interested in user-representing agents, the
most common way to describe agents’ reasoning
methods is through the Beliefs, Desires, and
Intentions – BDI – paradigm (Rao & Georgeff,
1995).
Trust is based on trust evidences (Melaye &
Demazeau, 2006), which are facts that are relevant
to the question of trusting an agent or not, which can
come from different sources. Common trust sources
are: direct experience, which can be positive or
negative, reputation, which is an evaluation that a
third-party agent provides about another one, and
systemic trust, which is the trust an agent has in a
group of other agents, without necessarily knowing
specifically each member of the group. These
evidences are stored in a way so that they can later
be used by a reasoning process, i.e. beliefs in the
case of BDI agents. Moreover, all trust knowledge is
contextual, i.e. it is related to an action or a goal Ω
the agent wants to perform or achieve.
Most commonly, these beliefs are split into a
small number of categories that are considered as
trust dimensions. The most used dimensions can be
described as ability, willingness and dependence
beliefs (Castelfranchi & Falcone, 1998). The belief
of ability means that an agent A believes that an
agent B is able to do what A wants it to do in the
context Ω. The belief of willingness means that A
believes that B will do what A wants it to do if A
asks it to do that. The belief of dependence means
that A believes that he relies on B to achieve its
goals in context Ω. There exist other dimensions, but
trust can often be easily simplified to retain only
these three ones without losing any accuracy.
Trust is then learnt through experience,
interaction and reputation transmission, stored as
agents’ beliefs and then used in decision-making
processes when agents have to make choices
involving other agents.
Amongst the large amount of existing trust
models, we need to rely on some criteria to make our
choice and ground our work on an adequate model.
2.1.1 Some Theoretical Criteria for Trust
Models
A trust model has to fulfill some criteria to be able to
be used in an agent network to create what is usually
called a trust network.
The first obvious criterion is optimization. This
comes from the fact that a trust model is made to
help agents to adapt their behavior to the network
configuration. So the trust model must be able to
improve the network’s global performance – the
ability for each agent to achieve its goals.
Optimization can only be tested experimentally,
because we are not able to foresee the network
improvement given a particular theoretical trust
model.
The second one is a practical requisite: the trust
model should be easily calculable, in order for the
agent to compute it in real-time without any
significant lack of reactivity; we call it calculability.
The third one is related to the fact that the agents
we describe are related to users. In many of these
systems, users often want to be able to understand
how the entity that represents them reacts. So, the
intelligibility criterion describes the ability of the
reasoning process and the semantic of stored beliefs
to be explained to the user and understood by him.
The other criteria we take care about are four
properties of the trust values (Melaye & Demazeau,
2006): observability, understandability, handlability
and social exploitability. They describe the ability of
the agent to apprehend other agents’ mutual trust, to
compute multi-dimensional trust values, to combine
these values into a global trust level and to use this
trust level to make decisions.
We will use these criteria altogether for both the
choice of the trust model to ground our integration
work and, later, for the integration model itself.
PERSONALIZATION OF A TRUST NETWORK
409

2.1.2 Falcone & Castelfranchi Trust Model
The trust model we used is the one introduced by
Falcone et al. (Castelfranchi & Falcone, 2004). It is
a BDI-based model – which corresponds perfectly to
the social network requisites – and uses numerical
representation for trust beliefs. Numerical
representations are better for a trust network than
logical representations, because these latest rely
quite always on complicated and high-complexity
modal logics, and so cannot be implemented.
This model uses three contextual values to
represent trust dimensions (cf. figure 1): the Degree
of Ability (DoA
Y,
), the Degree of Willingness
(DoW
Y,
) and the Environment Reliability (e()).
The latest is the only one that does not depend on the
agent B that is considered by A for the trust
evaluation. It measures the intrinsic risk of failure
due to the environment. The Degree of Ability
measures the competence that B has to accomplish
the task or to help A to accomplish the task Ω. The
Degree of Willingness measures the will of B to help
A to achieve its goals. All these three values are
defined in [0;1] and are combined into a global
Degree of Trust (DoT
Y,
) that describes the trust that
A has in B in context Ω through a function F. This
function is not specified in the model, but has to
preserve monotonicity and to range in [0;1].
DoT
Y,
= F(DoA
Y,
, DoW
Y,
, e()) (1)
Both of the ability and willingness beliefs – DoA
and DoW values – can be learnt from any trust
source. To learn these values, the agent uses a
reinforcement learning process that uses new trust
evidences to review its knowledge about others.
Figure 1: Falcone and Castelfranchi Trust Model.
We chose this model because it satisfies very
well the criteria we have for a trust model. The fact
that trust is computed from values representing its
dimensions with a simple formula guarantees the
respect of observability, understandability,
handlability and intelligibility. Social exploitability
is also respected because trust values provide a
ranking for potential partners that can be used to
make a decision. And then, this model is calculable,
because it is based on simple mathematical formula
and numerical values.
2.2 Agents Personalization
The other fundamental notion in this work is
personalization. Personalization is the ability for an
agent acting on behalf of a user to acquire and to
learn his preferences, his centre of interests and to
use them during its decision-making process.
While preferences nature is quite domain-
related, preferences representation has some
universal methods and properties.
2.2.1 Preferences Representation Models
The notion of personalization handles a couple of
distinct concepts. It is both a way to represent users’
preferences, a way to learn them from the user and a
method to use them in various contexts to improve
the agent’s behavior to the user’s point of view.
There are two distinct ways to represent users’
preferences (Endriss, 2006). They can be
represented by a valuation function giving a note to
alternatives the agent is faced to – and called
cardinal preferences. They can also be described as
a binary relation between each two of the
alternatives – and called ordinal preferences.
There are many ways to represent these two
kinds of preferences. But the most known and useful
are probably the weighted conjunction of literals for
cardinal preferences and the prioritized goals for
ordinal preferences. These models describe a way to
store user preferences but also a way to use them in
the reasoning process by evaluating and choosing
one between several alternatives.
All these preferences representation models can
be combined with several well-known reinforcement
learning techniques (Gauch et al, 2007), which will
enable them to improve the precision of the user
profile (i.e. the set of all represented user's
preferences) and adjust it to the user’s real
preferences. The learning process can use an explicit
feedback, which can be, for instance, a form that the
agent presents to the user. It can use an implicit
feedback, which is the analysis of the user activity,
for instance, the user web history for web navigation
assistants. Or it can use hybrid feedback, which is a
combination of both (Montaner et al, 2003).
ICAART 2009 - International Conference on Agents and Artificial Intelligence
410

2.2.2 Some Criteria for Personalization
As for trust models, we proposed a set of important
criteria that a personalization model has to respect in
order to be useful in an agent network.
In our work, we have kept four usual criteria
about personalization models (Endriss, 2006) and
added a new one. Firstly, the expressive power is
defined as the amount of preferences structures that
the model is able to represent. Secondly,
succinctness is the amount of information about
these preferences which can be stored in a given
place. Then, we have to take care about elicitation,
which represent the ease with which a user can
formulate his preferences in the model’s
representation language. This is an important
criterion, especially when a user is able to see
directly his profile and to modify or correct it on his
own. And finally, as it is also the case for trust
model, the complexity of the model is important in
order to be able to be computed in real-time by
agents.
Since the preferences learning mechanism is a
dynamic process, we have to describe the ability of
the system to react to any change in the user’s
preferences. This is why we add another criterion,
reactivity, which measures how much time the
model takes to adapt the profile to a change in the
user's behavior it represents.
3 INTEGRATION WORK
Concepts of trust and personalization having been
studied in a network of agents, we will now see why
there is a necessity to find a way to integrate these
two reasoning processes into a single one.
3.1 Personalization Integration in a
Trust Network
As trust and personalization needs to coexist in a lot
of networks – agent-based social networks, B2B
applications and negotiating calendars for instance –
we obviously need a way to make them function
altogether.
3.1.1 The Necessity of Integration
We can first believe that simply putting both models
on the same agents will be enough to make the
network work well. But this cannot be true, because
both being reasoning processes that cost much time
and resources and that leads to conclusions, there
will be two main problems happening. The first one
is that the cost of both inferences will be high for an
agent. The second one, and most important, is that
the two inferences can lead to different and perhaps
incompatible conclusions. And the problem will be:
how to handle these two conclusions and act while
taking both into account.
So the way to solve this problem is to create a
single reasoning process for agents, that takes into
account all the knowledge they possess, about both
the network (other agents) and user’s preferences.
3.1.2 Related Work
The only work we have found in the literature that
tries to integrate an approach of personalization and
trust (Maximilien & Singh, 2005) proposes a model
of multi-criteria trust in which the user has some
control over the importance of each evaluation
following a particular criterion in the final trust
calculus. This is a very limited and particular sort of
personalization, and this approach is not applicable
for the kind of networks we are interested in, as the
preferences are related to the trust model itself and
not to the domain the agents are concerned with.
So, to the best of our knowledge, no work exists
that tackles the interaction of these two notions in an
agent in the way of providing a single reasoning
process that handles both notions.
3.2 Considered Agents and Network
To be able to explain clearly how the solution we
propose works, we first need to describe the agents
and the types of networks in which it will be applied.
3.2.1 Agents’ Architecture and Capacities
The agents are based on the BDI architecture. This
means that (i) they possess some beliefs about their
environment – including the users of the system –
and other agents and (ii) that they all have some
goals, called desires, which are states – personal or
of the environment – they wish to be true. In order to
make these goals true, (iii) they use plans to make
decisions that become intentions – things they plan
to do.
As every agent does not have all the ability
needed to achieve every one of its goals, it has to
cooperate with some other agents in the network. To
minimize the cost of this required cooperation and to
avoid losing time and resources asking wrong agents
for help – wrong agents are those which cannot help
or will not help – it uses some trust process to
determine which agents are the best partners for a
PERSONALIZATION OF A TRUST NETWORK
411

specific task by the mean of the previously described
Falcone and Castelfranchi trust model.
Every agent should also be related to a single
user, and should be able to stock and use a
preference profile related to this user. We will see
later how the agents are able to do that.
3.2.2 The Network Structure
The network is merely an evolving set of agents
which are able to communicate one with another
through a message protocol that enables them to
exchange data, requests, answers and perhaps beliefs
and plans.
When a new agent is added to the network, it
knows a few other agents – its neighborhood – and
can learn the knowledge of other agents by
interacting with them. It order to make agents able to
learn the existence of unknown other agents, we
must include in the network a mechanism that makes
agents who are not able to process an information or
answer a request forward this request to another
agent. This mechanism involves only agents that are
not concerned by a request or information but are
cooperative – they are ready to help the request
sender or to distribute information in the network.
When such an agent receives an irrelevant message
from its viewpoint it does not ignore it but forwards
it to one or several agents of its own neighborhood
which it considers as the most able to answer to this
message; this mechanism is called restricted
relaxation (Camps & Gleizes, 1995). The number of
times a message can be forwarded in the entire
network is obviously limited to avoid cycles and
thus network overload.
3.3 Integration Constraints
As for trust and personalization models, we propose
some theoretical criteria needed to be fulfilled by an
integration model.
The first criterion derives directly from the fact
that we want to integrate a personalization model in
a trust model: this operation needs to result in an
improvement of the correlation between the user’s
expectations and the agent’s behavior. Hence the
criterion of accuracy will be the measure of
proximity between user’s desires and agent’s
observable behavior and results.
Two other criteria will simply describe facts that
are linked with the operation we are trying to do.
The target correlation describes the fact that the
alternatives considered by the preferences profiles
are defined by the integration model we choose. In
other words, the personalization model must be able
to evaluate the kind of alternatives that the trust
model will require it to evaluate. On the other hand,
the type correlation criterion describes the fact that
the personalization model should give as a result of
an alternative evaluation a value of a type that is
useful to the trust model.
Finally, because both of the two models we are
trying to integrate are contextual models, we have to
be sure that the contexts defined in both of the
models are compatible – which means that they are
the same or at least one is a subdivision of the other.
We will call it the context compatibility.
3.4 Towards an Integration Model
Taking into account all the listed criteria, and basing
our work on the chosen trust model, we developed a
solution for an integration of a personalization
model into this trust model. This integration has
been done in order to create a reasoning process that
handles information about both user’s preferences
and other agents’ behavior.
3.4.1 Description of the Proposed
Integration Model
The solution we propose simply consists in defining
a new dimension of trust in an already multi-
dimensional trust model (cf. figure 2). Indeed, in
order to take into account the personalization
evaluations, we considered that it was quite the best
solution to keep the preferences representations and
learning methods as separate agent’s ability. We
made this choice because it seemed very difficult
and confusing – both for programmers and users – to
incorporate it to the trust model reasoning process.
So this process is only going to use the evaluator
from the personalization model to rank alternatives
between other agents’ behaviors or results. It is also
going to learn a new context-dependant, agent-
dependant belief that represents the proximity this
agent’s behaviors or results have with the user’s
expectations concerning a particular context.
This belief, defined as a numerical value ranging
in [0;1], as other trust values, is also going to be
learnt by reinforcement learning methods using
personalization model’s evaluators. This will be
called the Degree of Similarity DoS
Y,Ω
. Then, the
global trust computation will have to be redefined as
another monotonic function F’ that ranges in [0;1],
which takes as parameters not only the parameters of
the function F from the trust model but also the
newly defined Degree of Similarity :
ICAART 2009 - International Conference on Agents and Artificial Intelligence
412

Do
T
Y,
= F’(DoA
Y,
, DoW
Y,
, DoS
Y,
,
e())
(2)
Moreover, this new dimension of trust can be
learnt from every trust source that is available in the
trust model we use. However, to learn it from
reputation an agent will have to take into account the
similarity level between itself and the evaluator, as
they do not have the same user profiles.
Figure 2: Modified trust model with integrated
personalization model.
The constraints that this integration model makes
on the choice of a preferences representation model
can be explained through the type correlation and
target correlation criteria: the Falcone and
Castelfranchi trust model uses numerical values in
[0;1] as trust dimensions values. So, as our new
belief will also have to be represented as such a
numerical value, and because, for learning, the
evaluated elements will be the results of one
interaction with another agent, the personalization
model’s evaluation functions will have to evaluate a
single result and give a numerical value as an
evaluation. So the personalization model should be a
cardinal preferences model.
3.4.2 Criteria Applied to the Proposed
Model
As previously shown, the original trust model we
used fulfills all required criteria. So, as we have just
added a single trust dimension, with its own
meaning and its own learning methods, these criteria
will not be broken. The intelligibility will also be
respected, because the meaning of this dimension is
easily explainable to the user; it represents the
proximity between the real behavior of the target
agent and the theoretical behavior it should have,
taking into account the user’s preferences. In fact,
only the optimization criterion needs to be
experimentally tested.
The criteria related to the personalization model
could easily be satisfied, because the choice of the
model is quite free between all the cardinal
preferences models. They can be satisfied, for
instance, by choosing the weighed conjunctions of
literals model, which is a light, low-complexity and
powerful model which can handle every domain of
application and is perfectly compatible – if correctly
implemented – with the contexts of the trust model.
Amongst the integration criteria, the two
correlation criteria are easily respected, as seen
before. The context compatibility can also be
satisfied by correctly implementing the
personalization evaluators. So, only the accuracy
criterion should be experimentally tested.
4 TESTS AND EXPERIMENTS
In order to first experiment the integrated trust and
personalization model we proposed and then to
check the two criteria that we can only validate
experimentally – optimization and accuracy – we
implemented a simplified version of the model,
using the agent programming language and IDE Jack
(http://www.agent-software.com/shared/products/).
4.1 A Simplified Model
We firstly simplified the model to test only the two
experimental criteria we exposed – the goal of the
test was not to determine the efficiency of the
Falcone and Castelfranchi trust model nor of any
personalization model, but to experiment if the
integration model itself is viable.
The first simplification we decided was the
implementation of a mono-source trust model. The
only source that we considered was direct
experiment. So the only trust evidences that were
taken into account by the trust learning process were
answers given by other agents to the sent requests.
The second simplification we decided concerns
the environment; it was supposed to be sure – every
message reaches its addressee – and resourceful – if
an agent has both the ability and the willingness to
perform an action, then the action is performed. So
we considerer e()=1.
We also selected simple functions for trust
computation and for learning. F’ (for complete
PERSONALIZATION OF A TRUST NETWORK
413

model) and F (for trust only model) are defined as
simple multiplications between each dimension.
DoT
B,Ω
= F’(DoA
B,Ω
, DoW
B,Ω
, DoS
B,Ω
) =
DoA
B
,
Ω
x DoW
B
,
Ω
x DoS
B
,
Ω
(3)
DoT
B,Ω
= F(DoA
B,Ω
, DoW
B,Ω
) =
DoA
B,Ω
x DoW
B,Ω
(4)
Learning method is defined as a weighted mean
of current and new values with fixed ratios – let DoX
be DoA or DoW, and a,d∈ [0,1].
For positive trust evidence:
DoX(t+1) = DoX(t) + a * (1-DoX(t)) (5)
For negative trust evidence:
DoX(t+1) = DoX(t) - d * DoX(t) (6)
And for the similarity value – in case of positive
or negative trust evidences:
DoS
(t+1) = ( DoS (t) + mod
DoS
(B,Ω) ) / 2 (7)
Finally, we decided to use a static representation
of preferences described by a very simple user
profile that enables an agent to rate every result it
receives in [0;1]. We emulated a preferences model
in that way, because preferences dynamics was not
very important for these tests and, given the high
number of possible preferences representations, this
would not be significant anyway. So we faked a
preferences model that would have reached a stable
state by attributing a simple static profile to every
agent.
4.2 The Experimental Protocol
We experimented in a network of 100 homogeneous
agents able to possess 3 basic capacities A, B and C.
Each newly created agent randomly receives the
ability to use each one of the 3 capacities with a
certain probability – p = 0.6 for most of the tests.
All the agents are able to communicate through a
message protocol defined in the Jack interface, and
all use limited relaxation paradigm – with a
maximum of 5 successive relaxations for a message.
We used randomly generated initial
neighborhood for each agent with a probability of
knowing each other agents equal to 0.1.
Then, the process continues step by step. At each
step, a goal is generated for each agent, which
consists in using a random capability: A, B or C.
Obviously, when the agent does not possess this
particular capability, it has to cooperate with other
agents to achieve its goal.
To emulate preferences evaluation, each result of
a capability A, B or C, is a document, which is
assessable by the personalization model of any
agent, according to a simple user profile it possesses.
So, when an agent uses one of its capabilities or gets
a result from another agent, it is able to rate this
result according to its own preferences profile. We
thus defined mod
DoS
(B,Ω) as the average pertinence
of documents given to A by B as an answer to a
request from A.
4.3 Experimental Results
We evaluated the number of messages exchanged
between agents and the number of goals that were
not achieved by agents to validate the optimization
criterion. We also evaluated the average pertinence
of results for the accuracy criterion.
Each value is measured at step 1 and step 100 for
4 different networks: (i) a simple network without
any model, (ii) a network with the simplified trust
model – without personalization –, (iii) a network
with the trust and personalization integrated model,
and (iv) a network with a model that only takes into
account the personalization value. Then the results
between step 100 and 1 are compared to measure
improvement. Expected results (cf. table 1) are an
increase for pertinence and a decrease for the two
other values.
Table 1: Experimental results summary: evolution
between steps 1 and 100.
Number
of
messages
Number
of goal
failures
Average
pertinence
No trust nor
personalization
-3.6% +1.6% -3.3%
Trust model
only
-13.7% -77% +3.5%
Trust and
personalization
model
-21.5% -76%
+11.9%
Personalization
only
+34.3% +27%
+42.3%
The results fit with our expectations: the number
of messages and the number of failures decrease for
all networks where there exist trust models, and the
average pertinence of results increases significantly
in networks where personalization is taken into
account.
Complementary observations can be made; for
instance, while in much cases network optimization
is the same for trust only and for trust and
personalization networks, we can observe that when
the pertinence results are too often very low, the
network obtains worse results, because of the
unbalanced importance of the different trust
ICAART 2009 - International Conference on Agents and Artificial Intelligence
414

dimensions; we can notice that this effect should
probably be corrected by adjusting correctly the
global trust computation function.
But, globally, we can conclude from these
experiments that the two experimental criteria that
we had expressed are satisfied, and so, that our
model seems to be an adequate solution to the
problem we wanted to address.
5 CONCLUSIONS AND
PERSPECTIVES
In this work, we have explored the possibility of
associating trust and personalization paradigms in an
agent network. We have done this in order to give to
agents the ability to handle both the intrinsic
uncertainty of a partial-knowledge, evolving
network, and the also evolving requirements of a
user’s set of preferences. Indeed, agents would have
to face them both in a social network in which each
user has one or more agent to represent him.
Knowing that just putting together the two
reasoning methods leads to heavy problems of
optimization but also to problems for mixing the
results given by each one, we have looked for a
solution to integrate both notions in a single
reasoning process. We first gave some theoretical
criteria to choose every component of a global
agent’s reasoning method that could handle both
trust and personalization: the trust model, the
personalization model and the integration model. We
then proposed a complete solution that is acceptable
according to those criteria.
Our solution involves the Falcone and
Castelfranchi trust model, to which we added a new
trust dimension, that we called degree of similarity.
It also involves a cardinal preferences model such as
weighted conjunction of literals, which is used by
agents to evaluate results and alternatives and learn
the degree of similarity they have with other ones.
The obtained experimental results for
optimization and accuracy criteria seemed to
validate these criteria. That is why even if these
experimentations were done on a simplified version
of the trust and personalization integration model,
we can say that the solution we proposed seems to
be viable and to be applicable to the kind of
networks we described. This model was developed
in order to improve the behavior of agents in these
open, partial-knowledge and user centered networks,
and it seems to achieve this goal.
Future work on this solution is to test it with a
full and multi-source implementation with dynamic
personalization from real users. As the Falcone and
Castelfranchi model is very powerful and because of
the large scale of different cardinal preference
implementations that can fit in the theoretical criteria
of this solution, we can foresee very different
solutions for various domains and the need to find
the adequate personalization evaluation and trust
evaluation functions to each model.
REFERENCES
Camps, V., & Gleizes, M.-P., Principes et évaluation d'une
méthode d'auto-organisation. 3
èmes
Journées
Francophones IAD & SMA, pp. 337-348. St Baldoph,
1995.
Castelfranchi, C., & Falcone, R., Principles of trust for
mas : cognitive anatomy, social importance, and
quantification. 3rd Int. Conf. on Multi-Agent Systems,
ICMAS’98, pp. 72-79, Paris, 1998.
Castelfranchi, C., & Falcone, R. Trust dynamics : How
trust is influenced by direct experiences and by trust
itself. 3rd Int. J. Conf. on Autonomous Agents and
Multiagent Systems, AAMAS’04, 2. New-York, 2004.
Endriss, U., Preference Representation in Combinatorial
Domains. Institute for Logic, Language and
Computation, Univ. of Amsterdam, 2006.
Gauch, S., Speretta, M., Chandramouli, A., & Micarelli,
A., User Profiles for Personalized Information Access.
The Adaptive Web, pp. 54-89, 2007.
Maximilien, E. M., & Singh, M. P., Agent-Based Trust
Model Involving Multiple Qualities. 4th Int. J. Conf.
on Autonomous Agents and Multiagent Systems,
AAMAS’05, Utrecht, 2005.
Melaye, D., Demazeau, Y. & Bouron, Th. “Which
Adequate Trust Model for Trust Networks?”, 3
rd
IFIP
Conference on Artificial Intelligence Applications and
Innovations, AIAI’2006, eds, IFIP, Athens, 2006.
Montaner, M., López, B., & De La Rosa, J. L., A
Taxonomy of Recommender Agents on the Internet.
Artificial Intelligence Review , 19, pp. 285–330, 2003.
Rao, A., & Georgeff, M. BDI Agents : From theory to
practice, Tech. Rep. 56, Australian AI Institute,
Melbourne, 1995.
PERSONALIZATION OF A TRUST NETWORK
415