Discussion questions for the blog...

[Subbarao Kambhampati <rao@asu.edu>]

Here are some things you can think about (and put your comments on the
blog).

1. In the regression example given in the class, we found that there
   was a state {~clear(b); holding(A)} that was generated as a child
   by regression, but it is clearly dead. It would be a good idea if
   we can kill every such state on birth so the search is improved. How
   hard do you think such a strategy going to be? (Specifically, is it
   guaranteed to be a computational win?)

2. Given a domain with K state variables and A actions, we said that
the size of the space searched by progression is 2^K and that searched
by regression is 3^K.  What is the size of the space searched by the
partial order planner?


3. We called it the "partial order" planning algorithm. However, the
   example we did in the class actually had a total order. Does this
make sense?

4. I said that the partial order planner will work on two types of
   flaws--open conditions and unsafe links. When it picks a plan, it
has to consider all ways of resolving it (for open conditions, this
involves all possible steps that can be added + the possibility of
making it true from the initial state; for unsafe links it involves
promoting as well as demoting the threatening step).

 Does the order in which the flaws are resolved going to have any
effect on the completeness (in particular, for our example, should we
have considered supporting ~cl(b) first and supporting hand-empty
first as two different choices in the search tree?)


5.(slightly harder) If I have actions with "conditional
effects"--where the action in essencse says that I will give you P if
Q is true at the time I execute. Will such conditional effects affect
the way a conflict between a step and a causal commitment gets
handled?
  
 (think, for example, the action, drive car from home to school, which
not only brings the car to school, but also brings everything in the
car to school. Suppose you want to leave your laptop at home and bring
the AI book to school. Suppose further that in the beginning, the
laptop is in the car and the car is at home. How will you solve this
problem in terms of progression, regression, and partial order
planning?)


cheers
Rao
[Sep 14, 2006]