Learning Action Models for Planning: An Overview of the Hedlamp Project

Hedlamp is a UK EPSRC grant funded research project in which we aim to tackle challenges with knowledge engineering of automated planning techniques when applied to real applications. Normally, successful deployment of planning technology relies on groups of planning experts encoding detailed domain models and investing large amounts of time maintaining them. We are developing a high level, application-oriented knowledge engineering framework usable by application developers who want to experiment with the potential of AI Planning, while encoding a precise domain model of some valuable application area. We are developing tools and theory for the framework which support knowledge acquisition, validation and operationality of the domain model. In particular, this project aims to explore the opportunities of applying model translation, adaptation and reformulation techniques to improve the model’s quality, and that of the planning function of which it is a part. In this paper we outline the main areas that Hedlamp has addressed, and overview an automated knowledge acquisition technique that has been tested with real industrial process data. Introduction: Project Context Hedlamp (Huddersfield and Edinburgh: Learning Action Models for Planning) is a UK EPSRC joint grant funded research project, 2012-16, in which we in which we aim to tackle challenges with knowledge engineering of automated planning techniques. Our work is carries on from previous knowledge engineering research platforms such as GIPO (Simpson, Kitchin, and McCluskey 2007) and ItSimple (Vaquero et al. 2012), but is also heavily influenced by the EPSRC AIS Programme’s industrial setting1. Utilizing planning machinery relies on the expertise of groups of planning experts having to learn application expertise, encoding detailed domain models of the relevant part fo the application, and investing large amounts of time maintaining the model. We are developing theory and techniques to meet the challenge through involving the expert in the encoding of the knowledge, developing V & V techniques which identify and remove bugs at the domain modelling stage, developing https://www.epsrc.ac.uk/files/funding/calls/2011/autonomousand-intelligent-systems/ techniques to acquire the knowledge automatically, and optimising the translation of the domain model into a form acceptable by a planning engine. The tools are built around the key concepts of expert involvement and inspection, translation, reformulation, and machine learning, This has led to the development of a high level, application oriented knowledge representation language (AIS-DDL), with translators to the family of PDDL planner input languages, as well as the development and use of an hierarchical planner which inputs AIS-DDL. Given planning knowledge is notoriously hard to encode, we need effective ways of removing bugs and certifying its quality. The interface needs multiple ways to perform V & V to enable this. Following this rationale, our progress has largely covered the following areas: • manual coding of an AIS programme application, expert involvement and validation: the knowledge representation lanaguage AIS-DDL has been developed such that the expert user can both add to it and inspect it. This involves it being accessible on a networked platform called ”KEWI”, which includes documentation and (links to) source material for traceability (G.Wickler, L.Chrpa, T.L.McCluskey 2014a). • parsing and translation: components of AIS-DDL are checked, parsed and translated down to an appropriate PDDL level for input to a planning engine. The translator embodies consistency checking, and the operational form is run to help in validation, hence V & V processes are being addressed (G.Wickler, L.Chrpa, T.L.McCluskey 2014b). • reformulation and heuristic learning: given the translation is automated, the resulting PDDL is not optimised and could include inefficient components. In this case, problem transformation components are useful to input the translated PDDL and transform it into a more efficient representation, as well as deriving heuristics in the form of macros. The transformations preserve the initial model’s semantics and keep it within the same PDDL language. We have shown that these techniques are both domain and planner independent (L.Chrpa, M. Vallati, T.L.McCluskey 2014). • internal planner: For validation by operation, we have developed a hierarchical planner within KEWI. This has the advantage of operational heuristics embodied within AISDDL which correspond to known procedures or activities involving a collection of primitive actions. • machine learning: where other sources of data can be identified (apart from expert involvement and the use of expert documents), then it is possible to consider the use of machine learning techniques either in real time or batch mode. For example, work in the area of ATC used tracking records to perform machine learning in the form of theory revision on an existing domain model (McCluskey and West 2001). From initial acquisition and domain maintenance viewpoints, the use of training data can lead to the automated reconstruction or evolution of the domain model. A central element of the HedLamp research project is in its aim to develop procedures to automatically learn domain models for automated planning. In this report we will give an overview of the recent developments in adding acton model learning to the KEWI environment, in particular to benefit from data output from inductrial processes. First, we will survey related work along these lines. Learning Domain Models for Planning