O-Plan: The open Planning Architecture

O Plan is an AI planner based on previous experience with the Nonlin planner and its deriva tives Nonlin and other similar planning systems had limited control architectures and were only partially successful at limiting their search spaces O Plan is a design and implementation of a more exible system aimed at supporting planning research and development opening up new planning methods and supporting strong search control heuristics O Plan takes an engineering approach to the construction of an e cient domain independent planning system which includes a mixture of AI and numerical techniques from Operations Research The main contributions of the work are centred around the control of search within the O Plan planning framework and this paper outlines the search control heuristics employed within the planner These involve the use of condition typing time and resource constraints and domain constraints to allow knowledge about an application domain to be used to prune the search for a solution The paper also describes aspects of the O Plan user interface domain description language Task Formalism or tf and the domains to which O Plan has been applied History and Technical In uences O Plan was initially conceived as a project to provide an environment for speci cation generation interaction with and execution of activity plans There are three distinct components see gure the planner s workstation or user interface the plan generator and the execution monitoring sys tem The main e ort has been concentrated in the area of plan generation Plan generation is di cult and it is a classic example of a search problem in AI The main target of this research is therefore search space control The outputs of this study are a better understanding of the re quirements of planning methods improved heuristics and techniques for search space control and a demonstration system embodying these results in an appropriate framework and representational scheme The story of O Plan starts from a software engineering viewpoint namely how to build an open architecture for an AI planning project with the aim of incrementally developing a system resilient to change It was our aim at the start of the project to build a system where it was possible to experiment with and integrate developing ideas Further the nal system was to be tailorable to suit particular applications Section brie y describes the components of the system in which the overall controller and some of the internal knowledge sources can be customised or replaced by the end user O Plan is intended to be a domain independent general planning and control framework with the ability to embed detailed knowledge of the domain Of course exibility can never be achieved The user does have to live with some basic design features and planning philosophies The primary limitation being the least commitment approach taken by the system O Plan grew out of the experiences of other research into AI planning particularly with Nonlin and blackboard systems We have included a taxonomy of earlier planning systems gure which places O Plan in relation to the in uences on its design It is assumed that the reader is familiar with these works as the bibliography does not cover all of them see for an introduction to the literature of AI planning The main AI planning techniques which have been used or extended in O Plan are A hierarchical planning system which can produce plans as partial orders on actions as suggested by Sacerdoti though O Plan is exible concerning the order in which parts of the plan are expanded An agenda based control architecture in which each control cycle can post pending tasks dur ing plan generation These pending tasks are then picked up from the agenda and processed by appropriate handlers hearsay ii uses the term Knowledge Source for these handlers The notion of a plan state which is the data structure containing the emerging plan the aws remaining in it and the information used in building the plan This is similar to the work of McDermott Constraint posting and least commitment on object variables as seen in molgen Temporal and resource constraint handling shown to be valuable in realistic domains by Deviser has been extended to provide a powerful search space pruning method The algorithms for this are incremental versions of Operational Research methods O Plan has integrated ideas from or and AI in a coherent and constructive manner O Plan is derived from the earlier Nonlin planner from which we have taken and extended the ideas of Goal Structure Question Answering qa and typed preconditions We have maintained Nonlin s style of task description language Task Formalism or tf and extended it for O Plan As with most planning systems intended to operate in realistic domains control of the search and the management of con ict between competing actions has been the focus of the work The eventual aim of all such systems is to incorporate techniques which should scale up to tackle the expected complexities