Re: ISSUE-126 (Revisit Datatypes): A new proposal for the real <-> float <-> double conundrum

As I've mentioned, I've done some prospecting on his issue with  
people who would be users of this feature. They considered the  
counting/discreteness issue completely uninteresting "Why would  
anyone want to do that?".

Consider perhaps a nonstructural restriction disallowing counting on  
floats. I definitely not a fan of features no one will implement and  
they are, in any case, at risk of being removed automatically by the  
two implementations criterion.

But personally, as an advocate for the inclusion of these types, I'd  
rather have mincardinality of a float/double range that corresponds  
to nonzero measure real interval be always satisfiable.

There does need to be some attention paid to the out of range  
numbers. Lexical numbers that fall outside the float range are  
specced to "round" to +/- INF. So

(1) PropertyRange( a:prop
         DatatypeRestriction( xsd:float
             minExclusive "n1"^^xsd:float
             maxExclusive "n2"^^xsd:float
         )
     )
(2) n1 is a lexical that when interpreted as a real is FLOAT_MAX + 1
(3) n2 is a lexical that when interpreted as a real is FLOAT_MAX + 2

(4) ClassAssertion( MinCardinality( 2 a:prop rdfs:literal) a:i )

In this case I suspect the desired behavior would be to have it be  
unsatisfiable, but I could check.


-Alan

On Jul 4, 2008, at 10:53 AM, Bijan Parsia wrote:

>
> On 4 Jul 2008, at 15:18, Boris Motik wrote:
>
>> OK, so what is the case for keeping the interpretation finite?
>
> I believe I've already made the case. The case rests on two points:
> 	1) Semantically, they are finite.
> 	2) Implementations already have to deal with large finite  
> datatypes via user defined ranges on integers. Thus, this makes  
> nothing *in principle* worse.
>
> Now, having such a datatype built in raises an affordance toward  
> dangerous practice. But I'd rather handle that with best practices  
> and advice, rather than mucking with the semantics in absence of  
> extensive experience.
>
>> I really believe this is difficult to implement correctly,
>
> I've no doubt of that whatsoever. I've conceded it.
>
>> and I repeat my example (slightly modified).
>>
>> (1) PropertyRange( a:prop
>>         DatatypeRestriction( xsd:float
>>             minExclusive "n1"^^xsd:float
>>             maxExclusive "n2"^^xsd:float
>>         )
>>     )
>> (2) n1 is a constant that corresponds to the number 1 * 2^-149
>> (3) n2 is a constant that corresponds to the number 3 * 2^-149
>>
>> (4) ClassAssertion( MinCardinality( 2 a:prop rdfs:literal) a:i )
>>
>> This ontology is unsatisfiable: the range of a:prop contains only  
>> one object, but (4) requires existence of two different objects.
>> The difficulty in detecting this is that you need to count how  
>> many numbers are there between n1 and n2. How are you going to do  
>> that? The binary representation of floats is really cumbersome to  
>> deal with.
>
> Er...floats (and doubles) are inherently binary.
>
> I'm confused. The floats have (at least) three issues:
> 	1) Exactness of operations (irrelevant since we don't have equations)
> 	2) Finiteness of the set (fixed range and discrete)
> 	3) Difficulties with the representation (which I always associated  
> with 1)
>
> 2 always struck me as the big problem. I pointed out that we  
> already have this problem with integers and user defined types. I'm  
> confused by your new point.
>
> Please clarify whether it's 2 or 3 you are concerned with. If 2,  
> please explain how it's different from other finite type ranges.
>
>> I firmly believe that, if we stick to the continuous
>
> Discrete, I believe you mean.
>
>> implementation of floats, NO reasoner on this earth will implement  
>> this test
>> case correctly. I am open to others showing me wrong.
>>
>> Furthermore, I firmly believe that such inferences are irrelevant  
>> for practice.
>
> We have a continuous floating point type already (decimal). We can  
> always defined a fixed range (with < and > facets) and a length  
> (with the digit limiting thing), so we can always recover a float  
> like type from decimals.
>
> Given that we have a continuous floating point type already, the  
> reason for picking float can rest on two bits:
> 	The weird extra constants (e.g., NaN)
> 	The discreteness.
>
> These two things are not separate, of course. The discreteness  
> (combined with inexactness of representation) are a driver for the  
> constants.
>
>
>> Finally, I don't see a point in producing a spec for which we know  
>> nobody will implement correctly and that is irrelevant for
>> practice.
>
> I know neither of these things. I suspect it's likely. But I'd  
> rather guide people away from using floats than to distort their  
> meaning. As I said, I can imagine wanting to reason about floats as  
> floats (e.g., for analyzing certain computational processes). I  
> don't know if that would work out or not, frankly, but given that  
> we have ways to work around floats (e.g., 'use decimal, people!') I  
> don't see the point of mucking with the semantics.
>
>> The simple solution
>
> It's a simple solution, certainly.
>
>> is to make floats continuous and be done with it. Implementations  
>> become trivial and users will never notice the
>> difference anyway (partly because they won't care and partly  
>> because implementations will assume a continuous interpretation
>> anyway).
>
> I'm uncomfortable with this because it's based on some (plausible)  
> assumptions that I don't know are true. I'd rather encourage people  
> to use the more feasible types, or to use floats in non-dangerous  
> ways.
>
> If I were designing things from the ground up, I wouldn't have  
> included floats at all.
>
> Cheers,
> Bijan.
>
>

Received on Friday, 4 July 2008 15:09:05 UTC