Optimal temporal planning using the plangraph framework

The past few years have seen a rapid development in AI Planning and Scheduling. Many
algorithms and techniques have been studied and improved to deal with more complex and
difficult planning domains.
One such innovation was Graphplan, first developed by Blum and Furst in 1995 and soon
became one of the best approaches for optimal classical planning systems. Planning systems
that use Graphplan’s plangraph framework can find optimal plans for temporal planning
problems, in which actions have durations. However, these systems have had strict
assumptions on the preconditions and effects of actions, for instance, effects happen only
at the end of the execution. In addition, the algorithm used in the solution extraction phase
of these plangraph-based systems does not take full advantage of the information provided
by the expansion phase to prune irrelevant search branches early.
With the ambition to make temporal planning problems more realistic, the thesis proposes
an extension to the Planning Domain Definition Language (PDDL) 2.1 level 3, to allow
actions to have intermediate effects. Our optimal temporal planning system, CPPlanner,
is introduced as the first Graphplan-based optimal planner to handle the richer temporal
domains (i.e. actions can have intermediate effects). Futhermore, the planner applies “critical
paths” as a backbone for the search in the solution extraction phase, so that irrelevant
search branches are pruned early. This improves the performance even in more restricted
temporal planning domains.
In our experimental evaluation, CPPlanner outperforms two leading plangraph-based optimal
temporal planning systems, TGP and TPSYS, in almost all test cases. The state-of-theart
optimal planner CPT and latest temporal planning domains in the international planning
competition in 2004 and 2006 are also used in the experimental evaluation.