Information Technology Reference
In-Depth Information
11.14.9 Comparison with Other Existing Models and Conclusions
Many existing trust models are focused on reputation, including how trust propagates into
recommendation networks (Jonker and Treur, 1999) (Barber and Kim, 2000) (Jøsang and
Ismail, 2002). On the contrary, our model evaluates trust in terms of beliefs about the trustee's
features (ability, willingness, etc.); reputation is only one kind of source for building those
beliefs (other kinds of source are direct experience and reasoning). In the present experimental
setting there is not any reputational mechanism (that we could also simulate in the cognitive
modeling), so a comparison with these models is not appropriate.
There are some other approaches where trust is analyzed in terms of different parts; they
offer a more concrete possibility for comparison. For example, in (Marsh, 1994) trust is split
into: Basic Trust , General Trust in agents, Situational Trust in agents. Basic trust is the general
attitude of an agent to trust other agents; it could be related to our model if it is considered as
a general attitude to delegate tasks to other agents in the trust relationships; in the experiments
already illustrated, we did not consider the possibility of introducing agents with the inclination
to delegate to others or to do the task themselves. In any case, the setting can certainly include
these possibilities. General trust is more related to a generic attitude towards a certain other
agent; the more obvious candidate in our setting is willingness, even if the two concepts
overlap only partially. Situational trust is related to some specific circumstances (including
costs and utilities, that are not investigated here); there is a partial overlap with the concept of
ability, that represents how well an agent behaves with respect to a certain task. So, the model
presented in (Marsh, 1994) is, to a certain extent, comparable with our one; however, it lacks
any role for the environment (more in general for the external conditions) and it introduces into
trust the dimensions of costs and utility that in (Castelfranchi and Falcone, 1998), (Falcone
and Castelfranchi, 2001) are a successive step of the delegation process that is presented in a
simplified way.
Our experiments show that an accurate socio-cognitive model of trust allows agents in a
contract net to delegate their tasks in a successful way. In any case, for better testing the model it
is be necessary to realize a set of new experiments in which we even allow the cognitive trustors
to learn from experience. While the learning of the statistical trustor is undifferentiated, the
cognitive trustor is able to learn in different ways from different sources. In the model for
each trust feature there are (at least) four different sources: direct experience, categorization,
reasoning, reputation; each of them contributes in a different way. More, higher level strategies
can be acquired: for example, depending on the environment and the task difficulty (number
of hits) an optimal weight configuration for the FCMs can be learned.
Other directions of work could be to experiment with agents starting with some a priori
knowledge about other agent's stats with a percentage of error, and in which they can refine
this percent by analyzing how well they perform in the delegated tasks. This is a kind of
statistical, not specific learning. However, in order to learn in a more systematic way, an agent
has to discriminate each single stat. In order to do this, they could analyze the incoming reports
(e.g. how many times an agent tries a task for willingness; how many times they perform it
for ability; how many reports they send for controllability). The controllability stat introduces
an upper limit even to how many learning elements an agent can receive, so it becomes even
more critical.
Finally, we would like to suggest that a reputation and recommendation mechanism is
included, in order to add another trust dimension to the simulations. In this way we could
Search WWH ::




Custom Search