A Fuzzy Implementation for the Socio-Cognitive Approach to Trust - Trust Theory: A Socio-Cognitive and Computational Model - page 334

the second (referred to as the cognitive trustor ), based on a more complex set of factors that
have to be considered before trusting an agent; in particular, both a set of trustor features
and environmental features.
In all our experiments the cognitive trustors perform better than the statistical ones, both
from the point of view of global successes (number of Credits ) and stability of behavior (less
standard deviation in the simulations data). The cognitive strategy models the agents and
environment characteristics more effectively and it allows the resources to be allocated in a
highly accurate way. Introducing a changeable environment does not decrease performance,
providing that it is considered as a parameter; but even if it is not considered, the results are
largely better than with a statistical trustor.
The fact that an algorithm that knows the real processes implemented by the agents when
achieving tasks uses a simulation mechanism of these processes for selecting the best per-
formances is quite predictable. For this reason we have made new experiments introduc-
ing a significant amount of noise in the cognitive agent knowledge. The results show that
the performance of the cognitive agent remains better than the statistical one up to an er-
ror of 40%. So, the cognitive trustor is very accurate and stable under many experimental
conditions. On the contrary, even with a large amount of data from learning (the training
phase), the statistical strategy is not performing well. Moreover, if the learning is done in
a different environment, or if the environment is particularly negative, the results are even
worse.
With a low number of hits (e.g. three) the task is designed to privilege ability over will-
ingness; however, augmenting the number of hits, the relative relevance changes. A strong
environmental influence shifts the equilibrium, too: it modifies the ability scores which become
more variable and less reliable. Modifying the relative weight of those parameters (depending
on the situation) into the FCM of the cognitive trustor can lead to an even better performance.
In the real time experiments, when time is implicitly introduced as an additional performance
measure, a variant of the cognitive trustor ,the first trustful trustor , becomes interesting: it
maintains high task performance (measured by credits) with a limited amount of time lost.
Introducing costs into the experiment leads the agents to maximize another parameter, gain,
with respect to tasks achieved. It is not always the case that more tasks mean more gain,
because many agents who perform well are very costly; in fact the best strategy optimizes
gains but not the number of achieved tasks.
Introducing a monitoring strategy, with the possibility of retiring the delegation and reposting
the task, introduces an extra possibility for the agents, but also another difficulty, because each
re-post is costly. Considering explicitly the controllability as a parameter for trusting an
agent gives a significant advantage, because more controllable agents - especially in real
time - enable a more precise distribution of the tasks and even a recovery from wrong
choices.
Depending on the situation (e.g. with or without environment; with or without the possibility
of retiring the delegation) and from the goals (e.g. maximize tasks, time or gains) the possible
delegation strategies are many. In all cases, trust involves an explicit and elaborated evaluation
of the current scenario and of the involved components - and our results demonstrate that this
gives a significant advantage in terms of performance with respect to a mono-dimensional
statistical strategy.