- From: Drew McDermott <drew.mcdermott@yale.edu>
- Date: Mon, 6 Dec 2004 13:16:00 -0500
- To: public-sws-ig@w3.org
> [Manshan Lin] > 1) SHOP2 approach seems to use a partially ordered set of tasks to > describe the user's requirements. I don't think it's a feasible way > comparing with goal formula. Actually, the underlying intention of > the abstract tasks are their outputs and effects, which are goal > formulas. However, these goal formulas may be achieved by different > abstract tasks. I don't understand your point. > 2) The HTN planning domain includes a set of operators (that is, > atomic service) and a set of methods (that is, composite service). I > don't know whether the process model of OWL-S intents to be shared > among agents. But from the viewpoint of security, I thow doubt on > the agents' willing to expose their process model, which may reflect > their business logic. The services don't expose their process models. They expose a client's _view_ of the process model. That is, the steps of a composite process are actions for the client to perform that will (in the normal case) achieve the client's goals. > 3) In [3], it extends HTN planning algorithm to handle > incomplete-information planning problem. While some information can > be provided by web services, some must be provided by the user > himself since we can't require the user to provide all the needed > information to accomplish the plan. The question is, how to > distinguish what information should be provided by other web > services and what should be provided by the user? As discussed in > the mailing list before, since we can't control other services, the > information gathered during planning may be different when > executing. As mention at the end of the paper, I agree that > inserting queries to the plan (leading conditional plans) is more > proper. I tend to agree that, in principle, information gathering should be a planned activity, not an activity done before planning (and therefore organized in a different way). However, there is a caveat which shrinks the difference between the two approaches. It is in general impractical, counterproductive, and (often) impossible to produce a conditional plan that anticipates all the branches that might occur as a result of gathering information. Hence a planning algorithm must be able to produce an incomplete plan, in which some --- maybe all --- branches end with the note "If you reach this point, plan some more." So planning, execution, and information gathering are inevitably interleaved. I should mention that no one, as far as I know, has a good theory of when to stop elaborating conditional branches. -- -- Drew McDermott Yale University Computer Science Department Some see the glass as half empty, some as half full. Surely the truth lies somewhere in between.
Received on Monday, 6 December 2004 18:16:01 UTC