- From: Drew McDermott <drew.mcdermott@yale.edu>
- Date: Mon, 2 Feb 2004 15:10:48 -0500 (EST)
- To: public-sws-ig@w3.org
> [Yuji Sakata] > I think AI has achieved great success in planning. But it looks like > practical use of AI planning is not spread as a method for composing > program or component. (Sorry if just a shortage of my knowledge) AI has in a sense achieved "great success." There are several algorithms developed in the last decade that produce much bigger plans from bigger domain descriptions than was possible before. However, these domains are quite artificial. They suffer from the usual classical-planning limitations that Jim Hendler has outlined, but also from severe weakness in what one might call "diagrammatic" as opposed to "logical" reasoning. A key feature of all such algorithms is the ability to reason backward from a goal to actions that could achieve it, and their preconditions, which become subgoals. This "regression" process is easy for actions described using Strips (or its descendant PDDL), but much harder for actions described as actions on objects described geometrically. Most realistic domains have involved some amount of geometry, so that AI planning has been of little use > Experts in Semantic Web Services ML look like conclude that AI planning > will be useful for composing web services. > > Please teach me what are the characteristics of web services which > practically enable AI planning to compose a service process, though it > is fairy difficult to compose programs or components to achieve a > requirement. The beauty of web services is that it is one of the few practical domains in which geometry really doesn't matter. Numbers do, but ordinary arithmetic reasoning can be assimilated into PDDL without many problems. Unfortunately, while AI planning benefits from this feature of the web-service problem, it has to overcome all those other limitations (no perfect knowledge, and the need for contingent and partial planning). So once again AI planning is still mainly in the theory stage, but at least it hasn't failed yet. > In my humble opinion, > 1. higher abstraction of the functionality of service than primitive > codes or components. > 2. useful standards are fixed like XML, SOAP, WSDL,RDF..., though > previous program language are components are both proprietary. > 3. 'just' increasing actual requirements for dynamic service composing > in the intra/internet, which are motivated from a decentralized > services environment. There was no need for run-time composition and was > just a need for design-time composition. > > I think "1" is the most important characteristic. I'm not sure I understand what 1, 2, and 3 are. -- -- Drew McDermott Yale University CS Dept.
Received on Monday, 2 February 2004 15:11:35 UTC