From: earlcox@earlcoxreports.com
Bill,
There is no disagreement here. In my modeling system, I incorporate two
features. First a collection of rules in a container (your Block) called a
Policy. The rules in a policy are run in parallel, the results from one
policy can be used in the antecedents of rules in another policy. I also
have a blackboard server that provides control over the scheduling and
execution of policies (allowing the quiescence and synchronization of
multiple policies). In my previous comments I hadn't intended to go into a
complete and comprehensive discussion of ways to build feed-forward fuzzy
systems and of course, I started with "not necessarily" meaning that there
are cases where the parallelism of fuzzy systems can be bypassed. And, of
course, it's not my objective here to educate the news group members in the
formal (and informal) methodologies used to build effective fuzzy models
(after four books and close to 100 articles in AI expert and PC/AI Magazine,
I am a bit exhausted from the effort! )
Big E.
"William Siler" wrote in message
news:49b9df3d.0401091620.51edb6bc@posting.google.com...
> "EarlCox" wrote in message
news:...
>
> > "Greg Chien" wrote in message
> > news:ggALb.6351$na.6640@attbi_s04...
> > > "Rich Shepard" wrote
> > > > GPearson wrote:
> > > >
> > > > > Is it valid to use hedges on consequent sets?
> > > >
> > > > Yes.
> > > > IF Weight is high and Height is very tall THEN
> > > > body_mass_index is slightly elevated.
> > >
> > > Besides, when the consequent sets use the same representation as the
> > > antecedent sets, they can be taken as antecedents to subsequent rules
for
> > > processing.
>
> > Not necessarily. You must remember that a fuzzy system is, in effect, a
> > parallel processing system. All the rules are essentially run in
parallel as
> > they contribute to each of the outcome fuzzy sets. When all the rules ha
ve
> > been run, then the outcome sets are defuzzified to find a value. Even if
we
> > do not want to defuzzify or if we are using second order fuzzy sets, we
> > still do not have a complete outcome representation until all the rules
have
> > been executed. This means that we cannot usually say, in the same
knowledge
> > base rule set,
> >
> > if A is High and B is Low then C is Elevated;
> >
> > if C is Elevated then D is Small;
> >
> > because C, as a variable, does not actually have a value until all the
rules
> > are run and we instantiate C with the scalar representation of the
> > underlying fuzzy set. The way a fuzzy rule-based is executed is
> > significantly different from conventional backward or forward chaining
> > techniques.
> >
> > In a fuzzy rule-based mode, each solution variable is invisibly
represented
> > by an "under generation" fuzzy set that is constructed from the outcomes
of
> > the rules that specify that variable in their consequents. The way the
> > underlying outcome fuzzy set is formed depends on the implication
method,
> > but, in general is it is based on the amount of evidence present in the
each
> > of the contributing rule predicates).
> >
>
> This is one of the very, very few times I disagree with Earl.
>
> True, in a single rule-firing step all fireable rules must fire before
> that step is completed. But there can be, and in my experience usually
> are, blocks of rules that subsequently fire in parallel, and blocks of
> rules after that, and so on. It is very common in my programs to have
> the consequent datum in one rule appear in the antecedent of a rule in
> the next block of rules to be fired. In fact, if one uses recursion,
> the same rules may be fireable several times in a row before they quit
> and make another block of rules fireable, as in solving a differential
> equation.
>
> William Siler
--- SoupGate-Win32 v1.05
* Origin: you cannot sedate... all the things you hate (1:229/2)
|
