Projection of a task onto a subset of variables (a “pattern”); the
abstraction underlying pattern databases.
Derived variables are dropped from the pattern, so conditions on them are
treated as satisfiable. A conditional effect whose conditions mention
variables outside the pattern may or may not fire, and both outcomes get a
transition. Both keep the abstraction admissible.
An operator gets one transition per abstract state matching its
precondition, times the product, over the pattern variables it may set
through such uncertain effects, of the number of values each may end with.
That is never more than the number of abstract states, but on a large
pattern the transition list can grow toward
operators × abstract states × abstract states; size patterns accordingly.