Information Technology Reference
In-Depth Information
Keeping them separate can lead to building different graphs and applying different revision
heuristics, too.
The process of evaluation of the sources is described as a step that follows the stabilization
of the FCM. In computational terms, thanks to the characteristics of the FCMs, these feedback
processes can even be made in a parallel way. We see this process as a cognitive updating of
beliefs: the parallel options are better to model gradual opinion shifts, while the two phases
division allows the mimicry of counterfactual phenomena and drastic after decision changes.
When is it good for the system to evaluate its sources? In our experiments we assume, in
a conservative way, that the mechanism only starts after a decision; the rationale is that a
decision taken is a sufficient condition to assume that the stabilization reached is sufficient;
less conservative criteria are possible, of course: this choice can be considered an heuristic
rather than a part in the way the system works.
An interesting 'side-effect' of this source evaluation mechanism is that revision has less
effect on well established sources (i.e. agents that have many interactions with us); we are
less inclined to revise the stronger ones, mainly for economic reasons. The process described
takes into account all the past experiences, so introducing a new example (or a counterexample)
has less impact on the case of many interactions.
11.14 Contract Nets for Evaluating Agent Trustworthiness
6
In this paragraph we show a first significant set of results coming from a comparison among
different strategies for trusting other agents using a contract net protocol (Smith, 1980). We
introduced three classes of trustors: a
random trustor
,a
statistical trustor
,a
cognitive trustor
.
All the simulations were performed and analyzed using the cognitive architecture AKIRA
(Pezzullo & Calvi, 2004).
The results show the relevance of using a
cognitive representation
for a correct trust attri-
bution. In fact, a cognitive trustor performs better than a statistical trustor even when it has
only an approximate knowledge of the other agents' properties.
11.14.1 Experimental Setting
We implemented a contract net with a number of trustors who delegate and perform tasks in
a variable environment. Each agent has to achieve a
set of tasks
and is defined by a set of
features:
ability set
,
willingness
,
delegation strategy
.
The
Task set
contains the tasks an agent has to achieve; it is able either to directly perform
these tasks, or to delegate them to some other agent.
The
Ability set
contains the information about the agent's skills for the different tasks: each
agent has a single ability value for each possible task; it is a real number that ranges in
(0,
the expressiveness of the explicit one. Sometimes the trade off between computational power and expressiveness can
lead to the adoption of the implicit mechanism.
6
We would like to thank Giovanni Pezzulo and Gianguglielmo Calvi for their relevant contribution to the imple-
mentation and analysis of the model discussed in this paragraph.
Search WWH ::
Custom Search