For the logistics domain we
used the test suite of the Aips-00
competition. The results are shown in Figure 14. In this domain Grt, as well as Ff and Stan,
performed well, solving all the problems. Hsp
and Altalt failed to solve the
large problems within the time-limit. In general, best plans are found by Stan, which uses special
domain-dependent heuristics for problems identified as transportation problems.
Best solution times are achieved by Ff
and Stan in the small problems
and by Grt in the large ones.
Figure 14: Solution length and time (msecs) for
the logistics problems of the Aips-00
competition.
The logistics
problems in Figure 14 have incomplete goal states. Grt ran with the most promising facts goals-completion
method and with the hill-climbing strategy. However, the incompleteness of the
goal state is an advantage for the planners that construct the heuristic in a
forward direction. Motivated by this remark, we forced all the planners to
solve logistics problems with complete goal states, requiring all the
trucks and planes to return to their initial location. The results are shown in
Figure 15.
Figure 15: Solution length and time (msecs) for
logistics problems with complete goal states.
In the new logistics
problems, Grt, Stan and Hsp-2 exhibited stable performance, solving the problems in
about the same time. For Grt,
this means that the goal completion mechanism behaves well, at least in this
domain. Ff failed to solve the
large problems. Finally, Altalt
solved some more problems and this is because the regression mechanism did not
encounter invalid states. Note that, although the goal state was complete in
this case, Grt treated these
problems as usual, attempting to enhance the goals.
Ioannis
Refanidis
14-8-2001