Optimal sequential planning is harder than satisficing planning.
While there is no difference in theoretical complexity
in the general case (Bylander 1994), many of the classical
planning domains are provably easy to solve sub-optimally,
but hard to solve optimally (Helmert 2003).
Moreover, strikingly different scaling behaviour of satisficing
and optimal planners has been observed in practice
(Hoffmann and Edelkamp 2005). In fact, this disparity
even extends to planning domains which are known to
be easy to solve optimally in theory. If we apply two stateof-
the-art optimal planning algorithms (Haslum et al. 2007;
Helmert, Haslum, and Hoffmann 2007) to the GRIPPER domain,
neither of them can optimally solve more than 8 of the
standard suite of 20 benchmarks within reasonable run-time
and memory limits, whereas the whole suite is solved in a
few milliseconds by satisficing planners like FF (Hoffmann
and Nebel 2001). Moreover, those 8 tasks are quickly solved
by breadth-first search, showing no significant advantage of
sophisticated heuristic methods over brute force.
Copyright
c 2008, Association for the Advancement of Artificial
Intelligence (www.aaai.org). All rights reserved