... domains¹

To solve a particular planning problem (i.e., construct a sequence of actions to transform an initial state to a goal state), planners require a domain theory and a problem description. The domain theory represents the abstract actions that can be executed in the environment; typically, the domain descriptions include variables that can be instantiated to specific objects or values. Multiple problems can be defined for each domain; problem descriptions require an initial state description, a goal state and an association with some domain.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... scheduling²

Scheduling is an area related to planning in which the actions are already known, but their sequence still needs to be determined. Flowshop scheduling is a type of manufacturing scheduling problem.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... time³

We used actual time on lightly loaded machines because occasionally a system would thrash due to inadequate memory resulting in little progress over considerable time.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... thereof⁴

We used the BUS system as the manager for running the planners [Howe, Dahlman, Hansen, Scheetz, von MayrhauserHowe et al.1999], which was implemented with the AIPS98 competition planners. This facilitated the running of so many different planners, but did somewhat bias what was included.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... ⁵

We thank Eugene Fink for code that translates PDDL to Prodigy.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... developer⁶

We decided against studying some of the planners in this way because the representations for their development problems were not PDDL.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... results⁷

One planner was the exception to this rule; in one case, the planner timed out far more frequently on non-development problems.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... failures⁸

We separated the two because we usually observed a significant difference in the distributions of time to succeed and time to fail - about half the planners were quick to succeed and slow to fail, the other half reversed the relationship.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... 1GB⁹

We propose this figure because it is the amount requested by some of the participants in the AIPS 2000 planning competition.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... planner¹⁰

Paul Cohen has advocated such an experimental methodology for all of artificial intelligence based on hypotheses, predictions and models in considerable detail; see Cohen (1991, 1995).

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.