Almost Perfect Heuristics
The performance of heuristic search is commonly measured
by the number of performed node expansions. Of course,
this measure depends on the search algorithm used; for example,
A (Hart, Nilsson, and Raphael 1968) will usually
explore fewer states than IDA (Korf 1985) in the same
search space, and never more (assuming that successors are
ordered in the same way).
Here, we consider lower bounds for node expansions of
the A algorithm with full duplicate elimination. Results
for this algorithm immediately apply to other search algorithms
that rely exclusively on node expansions and admissible
heuristic estimates to guide search, such as IDA, A
with partial expansion (Yoshizumi, Miura, and Ishida 2000),
breadth-first heuristic search (Zhou and Hansen 2006), and
many more. However, they do not apply to algorithms that
use additional information for state pruning, such as symmetry
reduction, and neither to algorithms that use fundamentally
different techniques to find optimal plans, such as
symbolic breadth-first search (Edelkamp and Helmert 2001)
or SAT planning (Kautz and Selman 1999).
How many nodes does A expand for a planning task