Abstrakt
The International Planning Competitions have led to development of a standard modeling framework for describing planning domains and problems – Planning Domain Description Language (PDDL). The majority of planning research is done around problems modeled in PDDL though there are only a few applications adopting PDDL.
The planning model of independent actions connected only via causal relations is very flexible, but it also makes plans less predictable (plans look different than expected by the users) and it is probably also one of the reasons of bad practical efficiency of current planners (“visibly” wrong plans are blindly explored by the planners). In this paper we argue that grouping actions into flexible sub-plans is a way to overcome the efficiency problems.
The idea is that instead of seeing actions as independent entities that are causally connected via preconditions and effects, we suggest using a form of finite state automaton (FSA) to describe the expected sequences of actions. Arcs in FSA are annotated by conditions guiding the planner to explore only “proper” paths in the automaton.
The second idea is composing primitive actions into meta-actions, which decreases the size of a FSA and makes planning much faster. The main motivation is to give users more control over the action sequencing with two primary goals: obtaining more predictable plans and improving efficiency of planning.
The presented ideas originate from solving the Petrobras logistic problem, where this technique outperformed classical planning models.