[PowerLoom Forum] forward-chaining issue

[Xingzhi Pan]

Thanks very much. It's clear now.

2005/12/1, Hans Chalupsky <hans at isi.edu>:
> Dear Pan Xingzhi,
>
> yes, the PowerLoom forward chainer only considers rules with simple
> antecedents and rules with complex antecedents that are marked as
> `:forward-only? TRUE'.  You can of course get the answer you are
> looking for by using backward inference via `ask':
>
> STELLA(163): (all-facts-of com1)
> (|P|(COMPANY COM1) |P|(NUMBER-OF-MANAGERS COM1 10) |P|(NUMBER-OF-EMPLOYEES COM1 50))
> STELLA(164): (ask (small-company com1))
> TRUE
> STELLA(165):
>
> The division of labor between forward and backward chaining is
> primarily a performance issue, since depending on fanout, some rules
> are much more efficiently run forward than backward and vice versa.
> You can help PowerLoom by annotating which rules you want to run
> forward and which ones backwards only.
>
> For example, if you want this to be added with the forward-chainer,
> you can mark the rule as "forward-only" with the `=>>' notation:
>
> STELLA(165): (assert
>                (forall ?c
>                  (=>> (and (number-of-employees ?c 50)
>                            (number-of-managers ?c 10))
>                       (small-company ?c))))
> |P|(FORALL (?c)
>    (=>> (AND (NUMBER-OF-EMPLOYEES ?c 50) (NUMBER-OF-MANAGERS ?c 10))
>        (SMALL-COMPANY ?c)))
>
> STELLA(166): (all-facts-of com1)
> (|P|(COMPANY COM1) |P|(NUMBER-OF-MANAGERS COM1 10) |P|(NUMBER-OF-EMPLOYEES COM1 50)
>  |P?|(SMALL-COMPANY COM1))
> STELLA(167):
>
> The proposition is only true in an inference-cache world associated
> with the current module, which is why it is printed with a question
> mark here.  These inference caches are blown away if there are any
> retractions or non-monotonic changes, since then they might not be
> valid anymore.
>
> Finally, you can use the PowerLoom classifier to eagerly infer all
> types of instances and add them to the KB:
>
> STELLA(177): (setq *classify-from-non-inferable-parents-only?* false)
> ()
> STELLA(178): (classify-instances pl-user true)
> u
> ()
> STELLA(179): (all-facts-of com1)
> (|P|(COMPANY COM1) |P|(NUMBER-OF-MANAGERS COM1 10) |P|(NUMBER-OF-EMPLOYEES COM1 50)
>  |P|(SMALL-COMPANY COM1))
> STELLA(180):
>
> Caveat: by default, the variable
> `*classify-from-non-inferable-parents-only?*' is set to TRUE which
> prevents COM1 to be classified (this is an optimization for
> classification of large hierarchies).  Currently, you can't change its
> setting with a PowerLoom command which is why I changed it with a Lisp
> statement above (or you would have to change the corresponding C++ or
> Java variable and then recompile).  In future versions we should make
> this available as a user feature that you can change.
>
> Hans
>
> >>>>> Xingzhi Pan <vengeance.storm at gmail.com> writes:
>
> > Hello, friends.
> > I've been trying forward-chaining in PowerLoom. It seemed that,
> > sometimes PowerLoom will add inferred facts into KB, and sometimes
> > not. Try to compare these:
>
> > (defconcept company)
> > (defrelation number-of-employees ((?c company)(?n INTEGER)) )
> > (defrelation number-of-managers ((?c company)(?n INTEGER)) )
> > (defconcept small-company ((?c company)))
> > (assert (forall ?c (=> (small-company ?c) (and(number-of-employees ?c
> > 50)(number-of-managers ?c 10))) ))
> > (assert (small-company com1))
> > (all-facts-of com1)
>
> > and
>
> > (defconcept company)
> > (defrelation number-of-employees ((?c company)(?n INTEGER)) )
> > (defrelation number-of-managers ((?c company)(?n INTEGER)) )
> > (defconcept small-company ((?c company)))
> > (assert (forall ?c (=> (and(number-of-employees ?c
> > 50)(number-of-managers ?c 10))(small-company ?c)) ))
> > (assert (company com1))
> > (assert (and (number-of-managers com1 10)(number-of-employees com1 50)) )
> > (all-facts-of com1)
>
> > I found that PowerLoom will do the inference only when the antecedent
> > is an 'atom' proposition. When it's something like (and P Q), nothing
> > seems happened. Have I done something wrong? Or it's designed so? Any
> > help would be appreciated.
> > --
> > Pan Xingzhi
> > _______________________________________________
> > powerloom-forum mailing list
> > powerloom-forum at isi.edu
> > http://mailman.isi.edu/mailman/listinfo/powerloom-forum
>


--
Pan Xingzhi