In this section, we present
comparative results between the Grt
planner and other planners. We decided to use Hsp-2
(Bonet & Geffner, 2001), Ff
(Hoffman & Nebel, 2001), Stan
(Long & Fox, 2000; Fox & Long, 2000, 2001) and Altalt (Nigenda, Nguyen & Kambhampati, 2000)[1]. All these planners took part in
the domain independent track of the Aips-00
planning competition. We selected these planners because Hsp-2 and Stan are state-of–the-art planning systems, Ff has been awarded for its outstanding
performance in the last competition and Altalt
is a new but very promising domain-independent state-space heuristic planner.
The aim of our
experiments is to have an overall view of the performance of the evaluated
systems. Performing pair wise comparisons between specific optimization
techniques is not possible, since these techniques are implemented on top of
different systems. Moreover, this kind of comparisons is out of the scope of
this paper, which focuses in the use of specific directions for constructing
the heuristic and traversing the space of the states, in the area of
domain-independent heuristic state-space planning, and not in the evaluation of
the numerous pre-processing optimization techniques. However, in the cases
where we identify the contribution of a specific feature in the performance of
a planner, we comment on this.
In order to have fair
comparisons, we used exactly the same problem and domain description files for
all planners. So, Grt ran without
XOR-constraints or numerical representation of resources. Moreover, although
the irrelevant object elimination technique is an integral feature of Grt, it had no contribution in these
domains, since there were not irrelevant objects. We believe that the absence
of irrelevant objects in these domains does not mean that this technique has
limited applicability, but it is an indication that more real domains for
testing purposes have to be used in the future, since the planning tasks in our
real-life are full of irrelevant objects. Finally, the domain enrichment
technique proved valuable for the elevator domain only. However, this
technique, as well as the goal enhancement one, has not to be seen as an
optimization technique, but as a way to overcome the problems that arise from
the backward direction of the heuristic construction.
We tested the planners
in several domains taken from the planning competitions and from the
literature, in the same workstation and within the 5 minutes time limit. The
results are presented in the following.
Ioannis
Refanidis
14-8-2001
[1] Stan is available at
http://www.dur.ac.uk/~dcs0www/research/stanstuff/stanpage.html
Ff is available at
http://www.informatik.uni-freiburg.de/~hoffmann/ff.html
Hsp-2 is available at
http://www.ldc.usb.ve/~hector/
Altalt is available at
http://rakaposhi.eas.asu.edu/altweb/altalt.html