Re: Rules WG -- draft charter -- NAF

[ Pat Hayes added on the Cc:, in case he wants to correct my memory of
what he agreed to, down in the last paragraph. ]

> >      { ?game a Game.
> >        NotKnown { ?game error ?e }.
> >      }
> >      =>
> >      { ?game a PerfectGame }
> >
> >But... Well, at very least that's non-monotonic.   Any conclusion you
> >reach like this MUST NOT be re-published, right?  Or, they must always
> >carry with them the warning "if we assume my knowledge of the world is
> >complete, then...."   And people wondering by and picking up this rule
> >... well, do you have any reason to think they would know where to
> >look for errors?   The naive user/software is almost certain to get
> >false results from NAF rules like this.
> 
> The approach taken in the declarative logic programs literature, and most 
> practical
> rule reasoning applications, solves this in a quite simple fashion.
> Every conclusion is drawn (and prescriptively sanctioned by the semantics)
> relative to a **specified set of premises***.   In your terms, that set of
> premises is "my knowledge of the world".

But in traditional rule systems, a solid characterization of that set
of premises is known by everyone who cares about the conclusions; in
my Semantic Web scenarios, at least, it's not.

> If a rule is written using NAF,
> then the rules which depend on NAF should be viewed as belief policies.
> KR is primarily about principles for reasoning with ***beliefs***, not 
> irrevocable truths.  (If you want irrevocable truth,
> stick to mathematics and religion not practical daily living and its 
> pragmatic reasoning!)

I don't see where the difference between belief and truth manifests
here.  I plan to try adopting your suggestion here and see how it
goes. 

> Your worries about dangers of usage are overblown.  There are decades of 
> successful applications usage
> of NAF.  Probably hundreds of thousands of users have written Prolog, for 
> example.

You're arguing that we'll be able to walk on water because many people
have walked before.  How many of those prolog programs would still
give only correst results if you randomly swapped clauses between
them?  1% 2% 0.01% On the Semantic Web, they all have to.  Asserting a
rule base which gives incorrect results when randomly combined with a
rule base someone else has asserted (and with some clauses randomly
dropped) makes one of you a liar.

> This principle of relativity to a premise set is essential to any 
> nonmonotonic KR, not just
> to NAF.

Indeed.  Pat Hayes is pretty clearly on record as saying Semantic Web
logics have to be monotonic.  And he's the logic guy who sat through
all the RDF Core discussions about what RDF was really about.  

As I recall, after some hours of arguing with him at the March 2003
DAML PI meeting (after your "rules debrief" presentation), you two
came to consensus that he was right in terms of the web at large, but
you were right that non-mon reasoning was okay for conclusions used
only *within* an agent.  Perhaps we need to revisit that agreement and
characterize that line more precisely, if we're to agree on where NAF
fits into the Rules charter.

And as my previous mail said -- it's not like we can't do NAF/non-mon
on the Semantic Web, we just have to explicitely name the premise
set.  I gave two examples of how that might be done with little pain.

    -- sandro

Received on Thursday, 13 November 2003 10:35:23 UTC