[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Comments on the project domains



Comments on the Project 1 submissions.

Folks:

I went through the projects and plan to return them tomorrow. Here are some common comments that came up:


Some common problems:


--It is important to have a clear idea of who the agent is who is doing the planning. Otherwise
the problem may seem sort of contrived because the planner is ascribing goals to multiple actions


--It is worth checking if the problem is "planning" oriented, "scheduling" oriented or "subgoaling" oriented.
Top-down theorem proving (of the sort you did in backward chaining reasoners) involves subgoaling
(i.e., to achieve P, achieve R and W. To achieve R achieve J etc). But it doesn't quite have subgoal interactions


Scheduling problems mostly deal with situations where the plans for individual activities are fairly template driven, and the
action is all about sequencing them over the resources


Planning problems will in general have both the above characteristics.




Things that more than one person found out:


--Inherited types would be a good thing to support in PDDL

--Planners don't have a notion of cost of the action

--FF wound up being the planner of choice in many cases mostly because it supported -ve preconditions and
ADL robustly (and not as often because of its speed)
Notice that supporting -ve preconditions in particular is not a big deal at all. Most planner writers are just
lazy and postpone it (and in the marketplace of planners, this could lead to loss of customers ;-)


--FF has an annoying habbit of saying "Goal can be simplified to True; empty plan solves it"--even in cases where
there are apparently non-trivial plans needed to solve the problem


--Durative actions and numeric quantities are needed in some cases.

--It is not possible to model creation and destruction of objects in PDDL
(leading to inelegant workarounds that require that all objects be declared upfront--
there was a beach domain where the student had to make all the sandwiches upfront :-)



What earned extra points:


--Domains that were very different from the standard planning benchmark domains
--Discussion on planner/PDDL features that helped/hurt in modeling the domain
--Discussion on what is hard about planing in the domain, and whether or not things had to be
abstracted out to make the domain doable for the current planners.
--Reports that looked into the type of plans generated and commented on whether or not the plans are the kind of
one that are best for that domain



[I will put up a couple of exemplary projects]


Some off-beat domains with good evaluations:

  "Aircraft battle damage repair domain"
   "Dragons, Damsels and Knights"
   "Database query planning as a planning problem"

off-beat domains (but need better evaluation)

   " Cracker:  planning to gain access to an arbitrary network of computers"
   " Software engineering project planning"

Interesting variations on faily known domains
   " CHEF--a (vegetarian only?) cooking domain"


That is it for now.


rao