Re: Planning under Description Logic ?--an obstacle towards WSAC

On Dec 24, 2004, at 1:04 PM, Manshan Lin wrote:
[snip]
> As mention in quite a few papers, in practice, the resulting plan of
> WSC is
> relative short, but the choices are vast since many concrete WSes share
> similar functionality. This fact shows that we are trying to find a
> simple
> plan in a large and unbounded planning domain.

As is usually the case.

> If we launch our planning from the initial state, we have to face a
> large
> number of irrelevant WSes.

Not necessarily. You could filter first. THere are tradeoffs.

>  Of cource, we can use some heuristics to compute
> the contribution of each operation, but there is too much cost in
> computing
> the irrelevant ones' heuristic.

Ok, at this point I start to lose interest. What is the *point* of this
armchair theorizing? Naive blind search sucks? Yeah, we know that :)

SHOP (and many HTN planners, but not all) works foward from an initial
state. It doesn't use *only* the initial state, and it is not aiming at
a goal. There is more in heaven and earth, Horatio.

And, as Drew pointed out, it's generally not the case that you just do
one or the other, or even something that is obviously characterized as
either.

Also, since once you get what many people to be the desired
expressiveness of your planning formalism, computational complexity
goes through the roof, there is a sense in which *all* your choices
suck, and you are going to have to figure out how to inject better
control strategies. Direction isn't sufficient, and one direction
usually isn't necessary. Optimizations and heuristics are key.

> Launching from "initial state" and launching
> from "goal" are not symmetric. What's more, most of the heuristics for
> planning
> (such as RePop, HSPR)aims at finding the shortest plan, not the best
> plan. I
> think the heuristics should be enhanced. The planning for WSC becomes
> more
> complicated when we take non-functional factors into consideration.
[snip]

Or less sine the non-functional factors might cut the domain
dramatically (consider "Geographic location").

Most importantly, and I speak as someone who has speculated in print on
this topic, we don't have a good sense of the shape of ``typical'' web
services domains such that we can usefully characterize what the space
is like. Alas. We all are reasoning a bit from our conclusions,
"backwardly", and in some ways its very backward indeed.

Cheers,
Bijan Parsia.

Received on Friday, 24 December 2004 10:37:02 UTC