- From: Bijan Parsia <bparsia@cs.man.ac.uk>
- Date: Tue, 27 Jan 2009 17:14:35 +0000
- To: Jim Hendler <hendler@cs.rpi.edu>
- Cc: W3C OWL Working Group <public-owl-wg@w3.org>
Just one small point. On 27 Jan 2009, at 15:47, Jim Hendler wrote: [snip] > 2 - there's a lot of question about EL vs DL - again, the primary > explanations are with respect to theoretical performance, but the > AC members aren't interested in that - as one pointed out to me > "there's a lot off polynomial stuff that doesn't work in practice, > and some exponential and undecidables that are in daily use" - he's > right. [snip] This is, of course, true. But there is fairly extensive experience supporting OWL EL, OWL RL, and OWL QL as robustly scalable. It's not just that they formally toe the line, but that they practically do so. (And even just theoretically, the *way* the have their complexity makes it pretty clear that they are robust...i.e., the likely sources of difficulty are pretty easy to see. For example, a naive OWL QL implementation can die if you have a huge, super deep hierarchy with a moderately sized query involving separate branches (because the query grows exponentially). This is not a likely case given that the likely application area or...actually, at all. Even there, it's pretty straightforward to deal with.) There are *at least* three independent implementations of EL++ (CEL, IBM, and in Pellet). All exhibit excellent scaling behavior on all inputs thrown at them (including SNOMED). Back in the day, when we were considering what profiles to include, I said that 2 *production quality* interoperable implementations should be the bar. We can define production quality in a number of ways, of course, and include scaling. I just what to emphasize that the the rationales are not with respect to *theoretical* performance (though that is part of the story) nor are the *primary* rationales. OWL QL is useless for SNOMED in spite of its theoretical better performance. Hence OWL EL. (Please don't respond that SNOMED is a narrow case. I just picked it because it's a clear example, not that it is the only example. It's large, significant, with fairly complex but intelligible modeling.) I'd be happy to chat with any AC members who are confused on this point and explain in detail why the performance of these profiles is not merely "theoretically" good. Cheers, Bijan.
Received on Tuesday, 27 January 2009 17:11:10 UTC