SEQ@dlgm.daresbury.ac.uk (DARESBURY SEQMAIL) (05/25/89)
---- Start of forwarded message. Via: UK.AC.DARESBURY.DLGM; Thu, 25 May 89 01:10 GMT (V40 at UK.AC.DARESBURY.DLGM) Via: UK.AC.NSFNET-RELAY; Thu, 25 May 89 00:56 GMT (V40 at UK.AC.DARESBURY.DLGM) Received:from net.bio.net by NSFnet-Relay.AC.UK via NSFnet with SMTP id aa15834; 24 May 89 23:38 BST Received:from nooksack.cs.washington.edu by net.bio.net (5.61/1.15) with SMTP id AA28865; Wed, 24 May 89 14:52:58 -0700 Received:by nooksack.cs.washington.edu (5.52.1/6.13) id AA02618; Wed, 24 May 89 14:51:29 PDT Date: Thu, 25 May 89 15:05 GMT From: Dan Weld <weld@edu.washington.cs.nooksack> Return-Path: <weld@nooksack.cs.washington.edu> Message-Id: <25 MAY 1989 15:05:41ooksack.cs.washington.edu> To: overt@com.unisys.prc.burdvax Cc: bio-matrix-arpanet@net.bio.net In-Reply-To: overt@prc.unisys.com's message of Wed, 24 May 89 15:02:11 -0400 <8905241902.AA13824@caesar> Subject: qualitative models in biology Sender: weld%edu.washington.cs.nooksack@net.bio.net 1. Computational tractability. Can you give an example of a qualitative reasoning system that has solved a serious problem for which a quantitative model exists but is computationally intractable? This one's not really implemented but gives one a feel for the rationale: Suppose you drop a cup of water, what's going to happen? Fluid hydronamics has a whole set of quantitative models that you might try to apply (Navier- Stokes equation, etc), but for the cocktail party domain, it is enough to know that the liquid will leave the cup, some will splash and some will spread along the floor. (Of course, it is nontrivial to create a reasonable qualitative model of this domain, as Hayes showed, but a qualitative model is clearly appropriate.) 2. > A quantitative model requires quantitative input. You can't use it if > you don't have it. This one I have no argument with and is one of the main reasons I think qualitative models are important in biology. But again, I see them as intermediates until quantitative models are feasible. For the example above, a quantitative model will never be feasible. What will you do when you drop the cup if you don't know the exact volume of liquid, its precise temperature (which influences viscosity), and the gravitational constant for your current location? Fluid systems are likely chaotic so very precise numbers are necessary here. Actually, what I really like are combined qualitative/quantitative models a la Erik Sandewall. This is interesting stuff, but has a number of problems. Basically, his paper suggests two things: integrating discrete and continuous representations (which I agree is long overdue) and abandoning discrete qualitative representations in exchange for quantitative analysis. This last point has several problems (which he admits). The most important is that it is not implemented and there is no way I can see to implement it. Nonlinear systems don't have closed form solutions. The earliest work on qualitative physics (de Kleer's SYN program) took exactly Sandewall's approach and tried to analyze electrical circuits by writing out Kirchoff's laws for all nodes and all loops and plugging them into the MACSYMA algebraic system. MACSYMA died. Yet novice EE students could analyze the circuits just fine. The trick is using qualitative reasoning when appropriate to focus the use of quantitative techniques and applying simplifying assumptions to abstract one's model. this is a huge area of research that I can't summarize here. 3. > Qualitative models support explanatation of reasoning so people can > understand why a program reached the conclusions it did. On the surface, I like the idea of keeping track of a chain of reaoning. However, I'm a more than a little suspicious that this feature will not prove useful when trying to model a truly complex system with lots of feedback loops (non-linear too), ie most of biology. It's true that this is a difficult problem, but there is some really interesting research on dynamic construction of abstract models. See my work on Aggregation (AI journal October '86) in a biochemical domain or the work on PROMPT by Addanki and the folks at IBM Yorktown applied to mechanical systems. In fact there is a whole chapter on this in our book, and a number of people are doing excellent research in this area. > Increased confidence. Here we part company. Qualitative models are very weak approximations to reality and so would seem to be the least likely to give correct predictions or explanations. How does someone even go about validating a qualitative model? There are fairly standard approaches for validating numerical models, but I haven't seen anyone address this issue for qualitative models. Any references? there certainly are some techniques. For example, Kuipers' has a soundness proof for QSIM (AI journal Sept 86) but no completeness result. In many respects the same approaches that work for standard simulation models work for qualitative models (and again qualitative models are not the only kind that are going to be of interest to Biologists); see my paper in this summer's qualitative physics workshop. And of course, this is an area for more research too. Dan ---- End of forwarded message.