On serializability definition..

[Subbarao Kambhampati <rao@parichaalak.eas.asu.edu>]

Okay granted--I might be overdoing this mailing thing a bit. But, then
I have been on a mail-diet for a whole week..


Consider the following scenario that seems to say that sussman anomaly 
is serializable.

Recall that in Sussman Anomaly, C is on A, A is on table, and B is
also on table. 

Suppose we consider the goal order On(B,C)-->On(A,B)

For On(B,C), suppose we make the plan:
 Put C on Table; Put B on C

Now for On(A,B),  We make the plan
 Put A on B

Now, we have both goals holding--and so the problem is
serializable. right?

But in class we said it is not serializable.. so what gives?

As Monika S realized and was bugging me at the end of the
class,  there is an implicit assumption in our serializability
definitions--that when we solve a subgoal, we will only consider
"non-minimal plans" for that subgoal. (The plan for On(B,C) above is
non-minimal since even if we remove the step of putting C on table, we 
are still getting On(B,C)).

In fact, it turns out that serializability definition is anything but
obvious. One can really only talk about serializability with respect
to a class of plans, and a class of planners. 

We may talk about this latter--but if  you are curious, check out

http://rakaposhi.eas.asu.edu/cse574/serial.pdf

G'night
Rao
[Feb  1, 2003]