Re: RM formalism supporting partial information
Date: Sun, 25 Nov 2007 23:49:34 -0400
Message-ID: <474a4254$0$5279$9a566e8b_at_news.aliant.net>
Marshall wrote:
> On Nov 23, 10:56 am, Jan Hidders <hidd..._at_gmail.com> wrote:
>
>>Exactly, so in that sense it is actually complete, and you can make >>that claim precise. The set of tupels in the answer will be exactly >>the set of tuples that are certain to be in the result of the same >>query over the omniscient database. By the nature of the problem every >>query should actually return 2 sets of tuples: the set of certain >>answers, and the set of possible answers. Your operators should >>therefore not operator on relations but on pairs of relations.
>
> It seems to me that anything that we can say about partial
> information can be said with total information. In other words,
> efforts at making the *system* understand partial information
> are merely pushing systemward calculations that could be done
> in a system without any understanding of partial information.
>
> If so, it seems to me the best we can hope for with such
> an effort is some additional convenience. At which point,
> any justification for a system with built-in support for
> partial information *must* be done in terms comparing
> the convenience of queries, processing, etc. with vs.
> without the new partial-info primitives. I don't recall having
> seen this done however.
>
> An analogous situation applies with approximate calculations.
>
> I would be interested to hear anyone agree or disagree.
Hmmmm... well, yes and no. Most of the discussions around here regarding "partial" information really center around missing information and various interpretations of null markers.
One could argue that in all the cases where those situations arise, the posited design is simply incorrect because it combines multiple relations using join when those relations do not have 1:1 relative cardinality.
One often sees examples where we know we have two employees Joe and Bill, and we know both of their salaries, and we know Joe is 47. I think the straw man constructing those examples is the combination of salary and age in the same relation.
If age were in a separate relation, one would have to explicitly state how to deal with Bill when age has any relevance. A simple join would explicitly state: "For all employees where we know the age..." by the simple act of joining the age relation, and this would exclude Bill. Other queries could include Bill but would explicitly state how to handle age.
As far as all that goes, I have to agree with you. I have to agree with you as far as probabilistic models go too. If we can assign probabilities, we can record what we know about those probabilities, but that is knowledge not partial knowledge.
On the other hand:
Science is fact based. Scientists make predictions based on what we know or have observed remaining silent on what we do not know or have not observed. Because mathematics has traditionally been used as a tool by scientists, it shows that same bias.
I am not sure what partial information would look like or how one would apply it. Someone someday might develop a truly useful concept of partial information and a truly useful theory of same. That person will no doubt be a lot smarter than I am, and his or her work will have revolutionary repercussions. If it ever happens. I won't go so far as to say it will never happen because I do not know that. I would not go so far as to say it will ever happen either because I do not know that either.
Thus far, null markers seem a barren sort of primrose path. I think the chance that any useful theory of partial information will arise from them is negligible. Received on Mon Nov 26 2007 - 04:49:34 CET