AI planning is a well-investigated framework for describing the evolution of system states through actions [6]. Action preconditions are first-order (FO) formulas, which are evaluated over states, i.e. finite sets of facts. Action effects either add or remove facts from the current state. Formulas are evaluated using a closed-domain, closed-world semantics, i.e. the domain is fixed a priori and facts that are not contained in a state are assumed to be false. The aim is to find a plan, i.e. a sequence of (grounded) actions that satisfy a goal formula. Knowledge representation formalisms are a natural way to introduce global constraints on permissible states; however, they usually interpret FO formulas over arbitrary models, instead of just one model with a fixed domain. DL-Lite Explicit-Input Knowledge and Action Bases (eKABs) [5] were proposed to combine classical planning with state axioms formulated in DL-Lite [4], while allowing an open domain and interpreting action preconditions under open-world semantics. Due to FO rewritability of DL-Lite, eKABs can be translated into the classical planning language PDDL. In theory, this allows the use of off-the-shelf planning systems. However, an initial evaluation [5, 13] using the Fast Downward (FD) planning platform [7] showed poor performance on a simple hand-crafted domain, with the planner being unable to solve even trivial problem instances. In our experiments, the problem also appeared in Fast Forward (FF) [9]. Planning Formalisms. A PDDL task is a tuple Π = hP, A, O, I, Gi with finite sets of predicate symbols P, action schemas A, objects O, the initial state I, and the goal G. I is a set of facts that are true initially; all facts not in I are considered false. G is a closed FO formula over P and O. Every a ∈ A is a triple hx−→a, prea, effai with parameters x−→a, precondition prea, and a set of effects effa. prea is an FO formula over P and O with free variables from x−→a. An effect e ∈ effa is a tuple h→−ye , conde, adde, delei where →−ye are variables, conde is the effect condition, an FO formula over P and O with variables from x−→a ∪ →−ye , the add list adde is a set of atoms with free variables from x−→a ∪ →−ye , and the delete list dele is a set of such negated atoms. For example, the action A = hx, Emp(x), h∅, >, {SoDev(x)}, ∅ii promotes any employee to a software developer.
A DL-Lite eKAB hP, A, O, T , I, Gi1 [
In [
The first pre-compilation (Ne), proposed by Nebel [
We therefore propose a second pre-compilation (DP) that avoids this overhead
by employing PDDL derived predicates [
As indicated by the O columns in Table 1, PDDL processing indeed constitutes the main bottleneck: In almost every unsolved instance, the planners failed already during PDDL processing. Part (b) shows that DP reduces the overhead of PDDL input handling and thereby increases overall performance in nearly every domain. TPSA and VTA are the only exceptions were the pre-compilation turned out detrimental. Besides an absence of complex conditions, in these domains the actions assign predicates to previously unbound objects. Therefore, by increasing the number of objects, the grounding and translation sizes grow drastically.
In contrast, the number of instances that could be solved by FF and FD after the Ne pre-compilation drops substantially in every domain. Apart from a blow-up of the file size, the Ne pre-compiler obfuscates the planning task’s original structure, which leads both planners’ searches into serious troubles.
In future work, we plan to design tools tailored to eKABs over DL-Lite and more expressive ontology languages, leveraging e.g. PDDL derived predicates. 2 https://gitlab.perspicuous-computing.science/m.steinmetz/pddl-dllite-benchmarks.git 3 https://gitlab.perspicuous-computing.science/a.kovtunova/moreflags2.git