PDDL and other languages..

PDDL and other languages.. : PDDL and other languages.. Lee McCluskey Department of Computing and Mathematical Sciences, The University of Huddersfield

Background : Background Related to AI Planning there are several kinds of knowledge that is required declaratively: Domain/environment Planning heuristics Task (input) Plans (output) – actions/activities Execution - resources

PDDL : PDDL A language convention for describing DOMAIN DYNAMICS (‘and nothing else’) which has succeeded through its use in the AIPS 98,00 and 02 planning competitions. Its aim of encouraging sharing of planning problems and algorithms has to some degree been achieved.

PDDL - form : PDDL - form PDDL’s syntax is LISP like. A domain (model) definition is structured into components by Keywords e.g. ‘:constants’ ‘:actions’ etc. The most important components are the set of actions. A special keyword is ‘:requirements’ which tells a process which blend of PDDL features are used in the domain definition. So we have a family of languages to suit planners with different capabilities.

PDDL - Semantics : PDDL - Semantics The basic requirement in PDDL is ‘:strips’ which indicates the underlying semantics of the language – worlds are considered as sets of situations (states), where each state is specified by stating a list of all predicates that are true. States are changed instantaneously into new states by actions which change the truth value of predicates. Actions have preconditions and effects under the default persistence assumption.. Etc…

PDDL - Where is the Semantics? : PDDL - Where is the Semantics? The semantics of PDDL v1.2 (used for ’98 competition) are informal and appear to be distributed among: the pre-existing languages/systems – strips, ucpop The v1.2 manual The language processors (solution checker) LISP somewhere else??

PDDL : PDDL Examples .. ‘The rule is that action definitions are not allowed to have effects that mention predicates that occur in the :implies field [RHS] of an axiom’ (p13) ‘An action definition must have an :effect or an :expansion but not both’ (p8)

PDDL - v2.1 : PDDL - v2.1 Extensions – different handling of numeric quantities, addition of durative actions Left out – HTN actions (apparently no-one had used them!) BUT – attempted to give a formal semantics to the language

Is PDDL a (good) modelling language?? : Is PDDL a (good) modelling language?? Fox and Long in the v2.1 manual describe it explicitly as one. Although not much discussed, PDDLv1.2 actually provides modelling features.. :timeless - predicates (static factual knowledge) :domain-axioms – written as L-R rules that form invariants on situations :expansion – allows encapsulation of actions in an HTN fashion :extends – allows some modularisation - one can import previously written components.

Is PDDL a modelling language?? : Is PDDL a modelling language?? But: PDDL was designed to reflect current languages and their underlying assumptions. It was NOT designed with a model building method in mind OR with many ‘pragmatic’ feature which make building easier. It is a ‘machine code’ rather than a language for human use!

Role of PDDL in the Semantic Web? : Role of PDDL in the Semantic Web? . One can imagine having planning services around the web – one supplies the problems + domain model in extended-PDDL and invokes the planning service. Extensions: Marked up (XML/RDF/RDFS (?)) version of PDDL Language for expressing advice / heuristics Service: analyses the domain model and configures a planner to solve the problems

Future – develop ontologies for Planning : Future – develop ontologies for Planning Ontologies are explicit specifications of a conceptual model for sharing the understanding of a particular domain. Some ontologies for planning concepts have been created – e.g. PLANET (Blythe) and SPAR (Tate). They are deemed essential for on-line agent communication between agents involved in planning (but promise multiple benefits eg in the KA process).

Future – develop ontologies for Planning : Future – develop ontologies for Planning Both - planning-oriented AND - planning-application-oriented ontologies need to be developed.

Future – my vision? : Future – my vision? Timely maturing of 4 research areas – Semantic Web KE – knowledge sharing and re-use Planning language conventions Planning KE Can be combined to solve the biggest problems in AI planning currently – lack of Accessibility and Usability of the technology