Information Technology Reference
In-Depth Information
11.5 Implementation with Nested FCMs
In order to understand the following parts of the work, we need to describe how a belief source
value is computed starting from many different opinions
3
of different sources (e.g. in a MAS
system, each agent is a source and can communicate an opinion about something). In an FCM
this situation is modelled with a set of nodes representing single beliefs that have an edge
ending on the (final) belief source node of the FCM. For each of these nodes, two values are
relevant: the value of the node itself and the value of the edge.
The value of the node, as usual in this kind of model, corresponds directly to the fuzzy label
of the belief; for example,
John says that the doctor is quite good at his work
can be considered
as a belief about the doctor's ability with value 0.5 that impacts over the others/reputation
belief source of a doctor's ability.
4
Computing the
impact factor
of this belief (i.e. the value of the edge in the FCM) is more
difficult. We claim that the impact represents not a cognitive primitive; rather, it has to be
computed by a
nested FCM
, that takes into account mainly epistemic elements about the
opinion itself and its source.
In our experiments with FCMs evaluation and epistemic issues are mixed up in a single
value; this was a methodological choice, because we wanted to obtain one single final value
for trustfulness. But this is the place where the two different kinds of information can be kept
separate because they have a different role. Figure 11.3 shows many elements involved in this
FCM: mainly beliefs about the source of the belief, grouped into three main epistemic features.
Here we give an example of such an FCM in the medical domain. This FCM has single beliefs
that impact on these features; the resulting value represents the final impact of a single belief
over the belief source node.
This nested FCM was filled in with many nodes in order to show the richness of the
elements that can intervene in the analysis. A similar FCM can be built for each single belief
that impacts into the belief sources nodes; some of those nested FCMs have overlapping nodes,
but in general each belief can have a different impact, depending on epistemic considerations.
It is possible to assign different impacts to the three different epistemic features; in this
case we wanted to give them the same importance, but it depends from both the contingent
situation, from personality factors and even from trust: for example, my own opinions can be
tuned by self-trust (e.g. sureness about my senses and my understanding ability), and Mary's
opinions can be tuned by trust about Mary. This leads to a very complex structure that involves
trust analysis about all the sources (and about the sources' opinions about the other sources).
For the sake of simplicity in the example we use all maximal values for impacts.
In general, it is important to notice that the 'flat' heuristic (same weights) we use in order
to mix the different factors is not a cognitive claim, but a need derived from simplicity and
lack of empirical data. In the following paragraph we investigate a very similar problem that
pertains to how to sum up the different belief sources.
3
We call this information
opinions
and not
beliefs
because they are not into the knowledge structure of an agent;
an agent can only have a belief about another agent's beliefs (
John says that the doctor is good
is a belief of mine, not
of John). This belief sharing process is mediated by opinions referred by John, but it can even be false, misleading
or misinterpreted. What is important, however, is that beliefs are in the agent's cognitive structure, whether they
correspond or not to other agent's opinions, beliefs or even to reality.
4
It is important to notice that this node does not represent an opinion of John; it represents a belief of the evaluator,
that can be very different from the original John's opinion (for example, it can derive from a misunderstanding).
Search WWH ::
Custom Search