Agre's Cooking problem

[From Jeff Hunter (920831)]

  [Re: Bill Powers (920827.0800)]

  I'm planning to get around to a lot of posts, but I
Have to comment on Agre's program (called Toast)
  It was tasked with making a breakfast, and did so (in
simulation) expeditiously, whereas a STRIPS planner blew up after
6 hours.
  This was (if you pardon the pun) a cooked example. None
of the subtasks interfered with any other in any major way.
  To show why, here is an example of weekend goals:

  - wear sunhat
  - get new car tires
  - spend weekend at cottage

  And these are fed into Agre's weekend planner (main mechanism
recapped below):

While there are unsatisfied goals,
   Choose aribtrarily an unsatisfied goal.

     ... use available tools and resources to satisfy goal.

  [pick a goal: spend weekend at cottage]
- walk to car
- enter car
- start driving to cottage
  [slack time. try other goals]
  [cannot wear hat. cannot walk to closet. am driving]
  [cannot drive to repair shop. am driving to cottage]
- arrive at cottage [goal #3 in progress]
  [pick a goal: wear sun hat]
- walk to closet [at home!]
- wear hat [goal #2 satisfied]
  [pick a goal: get tires changed]
- walk to car [at cottage!]
- drive to repair shop
- exit car. wait for tire change
  [slack time. try other goals]
  [pick a goal: spend weekend at cottage]
  [car unavailable. no problem]
- walk to cottage
- arrive at cottage [goal #3 in progress]
  [pick a goal: get tires changed]
- walk to car
- remove car from repair shop
  [pick a goal: spend weekend at cottage]
- drive to cottage
- enjoy weekend [soaking your feet]

  Since all goals require movement they all conflict. A bad
choice of goal ordering can make for a considerable amount of
walking!
  Any decent planner can find the best ordering for these
weekend goals in nearly the same time that Agre's one would
find a lousy one.
  The "planning weenies" have considered a lot of cases of
conflicting tasks, and their work should not be discarded lightly.

To insure that the breakfast maker problem was not entirely trivial,
we implemented a STRIPS formalization of a subset of the domain and
tested it using the SNLP non-linear planner [ref] on the omelette
making subproblem...
...(The planner ran for 6 hours on a Symbolics XL1200
before exhausting its paging disk). The planner was able to make
toast
and set the table however.

If this is what is meant by "too much machinery," I get the point. 6
hours!

  I think I see why the planner died. At least part of it
is that Toast is implicitly told that any pat of butter is the
same as any other pat, whereas the SNLP planner was not.
  Why are there 15 pats of butter? Why so that the planner
must consider more than 23 factorial orders of placing butter
and eggs in the frypan.

  Agre seems to have some good points, but don't swallow
his breakfast without a few grains of salt.
    ... Jeff

···

--
De apibus semper dubitandum est. Winni Ille Pu

[from Avery Andrews 920901.1312]
  (Jeff Hunter (920831)

On my reading, C&A's point is not that a certain amount of actual planning
is unnecessary, but that what planning there is gets done against a
background of routine activity, & lots of experience. E.g. mostly you
don't have to figure out the optimal order in which to carry out the
breakfast making subtasks, because you've done it lots of times before,
and saw people doing it before you had to. So the critique of the
planning weenies is not that planning is unnecessary, but that it would
be a good idea to find out more about the nature of the routine background
against which it happens. In fact, it seems to me that actual planning
by humans is incredibly bad when the background is deficient.

Avery.Andrews@anu.edu.au

I think Avery almost has it. Most formal planning starts
          with an attempt to fit a known template over a current set
          of facts (Pick A Model). Too many times this results in
          forced ignorance of precisly those things which make the
          current case unsuitable for analysis by the model chosen.
          "Best fit" means, very regularly, that certain facts and
          contingencies must be lopped off the decision tree.

          What is attractive about PCT for the planner is that one is
          forced to start with what is, and not what one must observe
          in order to find a best model fit.

(ps 920901.1600)

   [From Jeff Hunter (920831)]

     This was (if you pardon the pun) a cooked example. None
   of the subtasks interfered with any other in any major way.
     To show why, here is an example of weekend goals:

     - wear sunhat
     - get new car tires
     - spend weekend at cottage

     And these are fed into Agre's weekend planner (main mechanism
   recapped below):

i'm quite certain that neither agre nor horswill would dream of
suggesting using toast for planning a weekend. (and *i* certainly
wouldn't use a planner for my weekend!) is it unreasonable to assume
that different sorts of activities can use different sorts of planning,
in both the technical and the colloquial sense?

     Any decent planner can find the best ordering for these
   weekend goals in nearly the same time that Agre's one would
   find a lousy one.

substantiation?

     The "planning weenies" have considered a lot of cases of
   conflicting tasks, and their work should not be discarded lightly.

yes, the planning weenies have spent a lot of time worrying about
subgoal conflicts (sometimes called ``goal clobbers brother goal'').
one of agre's points is that usually you just don't have to worry
about it. and making breakfast is (claimed to be) a good example of
just that.

     I think I see why the planner died. At least part of it
   is that Toast is implicitly told that any pat of butter is the
   same as any other pat, whereas the SNLP planner was not.
     Why are there 15 pats of butter? Why so that the planner
   must consider more than 23 factorial orders of placing butter
   and eggs in the frypan.

have you ever seen a kitchen w/ one pat of butter in it? in fact i
bet you've never seen a kitchen w/ *any* pats of butter! i assume
there are fifteen pats of butter becasue that's roughly what you'd get
out of a stick and they didn't want to get into individuating the
stick into pats so they assumed it came that way. i think both toast
*and* the snlp would have a tough time simulating slicing off a piece
of butter!

   De apibus semper dubitandum est. Winni Ille Pu

;-}

        --Penni