Object-orientated planning domain engineering

Abstract

The development of domain independent planners focuses on the creation of generic problem solvers. These solvers are designed to solve problems that are declaratively described to them. In order to solve arbitrary problems, the planner must possess efficient and effective algorithms; however, an often overlooked requirement is the need for a complete and correct description of the problem domain. Currently, the most common domain description language is a prepositional logic, state-based language called STRIPS. This thesis develops a new object-orientated domain description language that addresses some of the common errors made in writing STRIPS domains. This new language also features powerful semantics that are shown to gready ease the description of certain domain features. A common criticism of domain independent planning is that the requirement of being domain independent necessarily precludes the exploitation of domain specific knowledge that would increase efficiency. One technique used to address this is to recognise patterns of behaviour in domains and abstract them out into a higher-level representations that are exploitable. These higher-level representations are called generic types. This thesis investigates the ways in which generic types can be used to assist the domain engineering process. A language is developed for describing the behavioural patterns of generic types and the ways in which they can be exploited. This opens a domain independent channel for domain specific knowledge to pass from the domain engineer to the planner.