Most planning problems are of hierachically structured, in which case
we are supposed to be able to retrieve a solution by doing causal analysis in a greedy fashion. However, simply applying this technique
assumes that all the goals are interleaving/disjoint, which is typically not the truth.
To take advantage of the structures, but in the meanwhile not be hampered by its limitation, we propose DTG-Plan.
It is an anytime algorithm planner consisted of two components. The higher level component uses the A* search algorithm which searches in an abstracted search space that expands each node during the search by one transition in a particular Domain Transition Grapha (DTG). The expanded transition might be of only one single action (if all preconditions are satisfied) or a sequnence of actions (if not all of the preconditions are satisfied). The corresponding sequence of actions for a transition is resolved by the lower level component, which is a greedy causal analysis based algorithm. It conducts causal analysis and retrieves a valid plan very fast. The upper level search component keeps increasing its scope (i.e. to a larger subset of the DTGs for the original problem), and incrementally increases the quality of the solution.
The parser for PDDL used in DTG-Plan is mostly based on the parser developed in Fast-Downward by Malte Helmert.