Re: Composition as planning

> [Jim Hendler]
>   easy example - suppose I call a service to see if Amazon has the 
> book I want in stock.  I then complete the plan and move to execution 
> - but while I was completing the plan someone else buys the last book 
> in stock -- my plan now fails (esp. if this book was to be used as a 
> precondition for something else) 

It's not clear what the import of this example is.  Yes, the plan
fails, but it was still a good plan.  The agent who used it must now
try again.  What else would you like to see occur?

> most AI planning 
> assumes all execution is under the control of the planner 

yes, more or less

> -- also 
> that operators execute instantly 

no; we have "durative actions" as of the last International Planning
Competition. 

> and without failure 

Yes in the sense that they don't do contingent planning.  But failure
is a special case, and perhaps not the most important, of contingent
planning.  If the planner plans to find out a piece of information,
and if the plan for achieving later goals depends on what that piece
of information is, then the planner has to plan for all the important
categories the information could belong to.  

> and lots of 
> things like that --- there are some planners that can handle certain 
> models of uncertainty and things like that, but none of thoe really 
> work that well for Web Services...

For web services, I think contingent planning is _much_ more important
than modeling uncertainty, or adversarial planning, or any of the
other variants that have been thought of.

-- 
                                             -- Drew McDermott
                                                Yale University CS Dept.

Received on Monday, 2 February 2004 14:38:59 UTC