The Use of Condition Types to Restrict Search in an AI Planner

Condition satisfaction in planning has received a great deal of experimental and formal atten tion A Truth Criterion lies at the heart of many planners and is critical to their capabilities and performance However there has been lit tle study of ways in which the search space of a planner incorporating such a Truth Criterion can be guided The aim of this document is to give a descrip tion of the use of condition type information to inform the search of an AI planner and to guide the production of answers by a planner s truth criterion algorithm The authors aim to promote discussion on the merits or otherwise of using such domain dependent condition type re strictions as a means to communicate valuable information from the domain writer to a general purpose domain independent planner Introduction to Condition Typing Research in AI planning has introduced a range of progressively more powerful techniques to address in creasingly more realistic applications Allen Hendler Tate A lesson learned in the expert systems and knowledge based systems eld is that it is impor tant to makemaximumuse of domain knowledge where it is available in order to address many real problems One powerful means of using domain knowledge to re strict and guide search in a planner is to recognise ex plicit precondition types as introduced into Interplan Tate and Nonlin Tate and subsequently used in other systems such as Deviser Vere Sipe Wilkins O Plan Currie Tate and O Plan Tate Drabble Kirby An explicit account of the Goal Structure or teleology of a plan can be kept in these systems This records O Plan work is supported by the us Advanced Re search Projects Agency arpa and the US Air Force Rome Laboratory acting through the Air Force O ce of Scienti c Research afsc under contract f c The project is monitored by Dr Northrup Fowler iii at Rome Laboratory the causal relationships between actions in the plan and can show the intentions of the domain writer or planner in satisfying conditions on actions In some circumstances such domain knowledge can be used to prune the search of a planner The information is pro vided to the planner via a planner s domain description language e g Task Formalism tf in Nonlin and O Plan The domain writer takes the responsibility for a deliberate pruning of the search space or for pro viding preferences via condition types This caused us to adopt the term knowledge based planning to describe our work Nonlin and O Plan tf extends the notion of a pre condition on an action and mates it with a process oriented view of action descriptions A tf schema de scription speci es a method by which some higher level action can be performed or higher level goal achieved Each schema is thought of as provided by its own manager The schema introduces lower level actions under the direction of its manager and uses that man ager s own resources The schema may say that some speci c sub action is included in order to set up for some later sub action as part of the overall task In tf such internally satis ed requirements in actions are speci ed as supervised conditions The man ager also relies on other normally external agents to perform tasks that are their own responsibilities but a ect the ability of this manager to do the task These are given as unsupervised conditions There are other conditions which the manager may wish to impose on the applicability of particular solutions e g don t try this method for house building if the building is over ve stories tall These are termed holds and usewhen conditions in di erent versions of Nonlin and are now called only use if conditions in O Plan Condition typing can be used to restrict search in a planner but there is work to be done on how far this technique can be developed It is often di cult for a domain writer to choose the correct type for a condi tion to most e ectively restrict the search space while not over indulging and throwing away plans which should be considered valid in the domain Tool sup port to aid in the reliable modelling of large domains will undoubtably be needed In practice we have found that condition typing is an essential aspect of encoding realistic problems to an AI planner in order to reduce search spaces to a manageable level