Pre-processing planning domains containing Language Axioms ∗ Marina Davidson and Max Garagnani Department of Computing, The Open University, Milton Keynes, MK7 6AA, U.K. {m.davidson, m.garagnani}@open.ac.uk Abstract We present an automatic procedure for pre-processing planning problems containing language axioms, a specific type of domain axioms. The axioms considered are assumed to be in the form p 1 ∧ p 2 ∧ …∧ p n → c. The pre-processing approach described consists of encoding the language axioms directly inside the given operators, contrary to other (not always correct) existing approaches in which axioms are converted into additional operators. A distinction is made between two different types of axiom sets, namely, recursive and non-recursive sets, which are treated differently. While a simple replacement algorithm is appropriate for the non- recursive case, a more sophisticated procedure is proposed for pre-processing sets of recursive language axioms. Subject to some assumptions, the method presented can be generalised and used with any set of axioms of the type considered. Preliminary experimental results seem to suggest that the approach identified is sound and may lead to improvements in planning performance. 1 Motivation One of the important issues in AI planning is how to define domains and problems in order to find a compromise between the expressiveness of a domain-definition language and the planners’ ability to handle it. More expressive languages are necessary in order to model the real world and allow a natural, easy and correct encoding of realistically large problems. However, expressive languages require more complex (and, in general, slower) planning machinery. The underlying motivation of this work is to allow more expressive domain-definition languages while simultaneously (1) retaining good performance and (2) avoiding (whenever possible) the redevelopment of existing planning technology, which is continuously required for bringing and maintaining up-to-date ‘old’ planners that cannot handle new, expressive languages. To achieve this, we propose the adoption of a pre-processing approach. The main idea behind pre-processed planning is to build automatic tools that translate domains (and problems) from expressive representation languages into simpler ones, for which fast and efficient planners already exist. This approach allows preserving requirements (1) and (2); in addition, it offers the advantages of modularity, simplicity and non-specificity (a pre-processing tool can be plugged into any planning system that accepts as input the chosen ‘target’ domain-definition language). In this paper, we focus on a specific feature of domain definition languages, namely, that of domain axioms. According to the description given in the original Planning Domain Description Language (PDDL) document, “axioms are logical formulas that assert relationships among propositions that hold within situations” (McDermott, Knoblock et al. 1998). A domain axiom a has the format context(a) → implies(a), where context(a) is a well-formed formula containing predicates of the language (possibly containing existential and/or universal quantification) and implies(a) is a one-predicate expression which is to be considered ‘True’ whenever context(a) is True. ∗ This work has been partially supported by the EPSRC, grant no. GR/R53432/01.