Mini Project 1: Hands-On Planning

You may work with one partner if you choose to.

---------------------------------------------------------------

In this project you will experiment with a planner of your choice. The
first step of this project is to pick a planner to experiment with.

Picking a Planner:
------------------

There are a number of publicly available classical planners, and you
are free to pick any one (or more) of them to experiment with. A good
choice would be the BlackBox planner as it has a reasonably well
maintained distribution. You can find BlackBox at,

http://www.cs.rochester.edu/u/kautz/satplan/blackbox/index.html

There you can download BlackBox in pre-compiled form (a number of
platforms) or the source. You can also download a number of example
problem domains and problems. The language that Blackbox accepts is
PDDL (a fancy way of specifying STRIPS schemas). The best way to
understand the language is to look at the provided examples. However,
for a formal description of the STRIPS subset of PDDL you can consult
the following document:

http://eecs.oregonstate.edu/ipc-learn/strips-pddl-subset.pdf

Note that there have been many extensions to PDDL that handle more advanced features than simple STRIPS (e.g. durative actions, concurrency, numbers, etc), but those features are not necessary for this assignment.

Note that BlackBox typically takes two PDDL files as input. One
specifies the domain operators, and the other specifies the initial
state and goal. 

Some other good choices of planners with available implementations are,

SATPlan 
http://www.cs.rochester.edu/u/kautz/satplan/index.htm

HSP (or HSPr) 
http://www.ldc.usb.ve/~bonet/#software

GraphPlan,
http://www-2.cs.cmu.edu/~avrim/graphplan.html

I don't have much experience with these distributions, but generally
they are quite straightforward to get running.

PART I: Playing with Blocks
---------------------------

Here you get to experience the blocks world first hand.

  1) There are many ways to define the blocks world. Create two
     versions of the blocks world, by writing two different sets of
     action schemas (the two sets may involve different types of
     facts). Try to design one set of operators such that they will be
     more difficult to plan with than the other. 

     You should provide me with the operator descriptions and a SHORT
     explanation about why you expect one set to be more difficult
     than the other.

  2) Create three blocks world problems involving 10 blocks each. Try
     to design a "very easy" problem, a "very hard" problem, and an
     intermediate problem. You will probably want to consider the way
     your planner works when judging the hardness of a problem
     instance. 

     You should provide me with a description of these three
     problems. Explain your intuition about the difficult of each.

  3) Use your planner with each set of operators to solve each of the
     problems (a total of 6 experiments) and record the basic
     statistics such as planning time, plan length, and whatever else
     may be of interest.

     You should provide me with the statistics and an assessment of
     how "good" the plans looked for each case. Were your predictions
     about hardness correct? 

     We will set a planning threshold of 5 minutes. So if your planner
     hasn't found a plan by that time it is considered a failure. If
     you find that the planner is failing on more than half of your
     problems, then you should decrease the number of blocks. 

PART II:
--------

Create your own planning domain. Be creative. The only constraints are
that it should easily incorporate parameters that control the hardness
of the planning problem (for example in the blocks world the number of
blocks is a hardness parameter). 

You should then evaluate your selected planner on your planning domain
by varying the problem difficulty. The details of the evaluation are
up to you. 

You should provide me with:
    
    1) A concise description of your problem domain.

    2) The action schemas for your problem domain.

    3) Some sample from your domain.

    4) A short evaluation of the planner's performance in your problem
       domain. Some questions that you might address are: The limits
       of the planner? Do the plans look "intelligent"? Why is the
       planner failing when it does fail? How does the planner scale
       with the problem difficulty? etc.

All of the material should be submitted in in electronic format by
email (afern@eecs.orst.edu). One project submission per team.