Knowledge engineering techniques for automated planning

Formulating knowledge for use in AI Planning engines is currently some-thing of an ad-hoc process, where the skills of knowledge engineers and the tools they use may significantly influence the quality of the resulting planning application. There is little in the way of guidelines or standard procedures, however, for knowledge engineers to use when formulating knowledge into planning domain languages such as PDDL. Also, there is little published research to inform engineers on which method and tools to use in order to effectively engineer a new planning domain model. This is of growing importance, as domain independent planning engines are now being used in a wide range of applications, with the consequence that op-erational problem encodings and domain models have to be developed in a standard language. In particular, at the difficult stage of domain knowledge formulation, changing a statement of the requirements into something for-mal - a PDDL domain model - is still somewhat of an ad hoc process, usually conducted by a team of AI experts using text editors. On the other hand, the use of tools such as itSIMPLE or GIPO, with which experts gen-erate a high level diagrammatic description and automatically generate the domain model, have not yet been proven to be more effective than hand coding.

The major contribution of this thesis is the evaluation of knowledge en-gineering tools and techniques involved in the formulation of knowledge. To support this, we introduce and encode a new planning domain called Road Traffic Accidents (RTA), and discuss a set of requirements that we have derived, in consultation with stakeholders and analysis of accident management manuals, for the planning part of the management task. We then use and evaluate three separate strategies for knowledge formulation, encoding domain models from a textual, structural description of require-ments using (i) the traditional method of a PDDL expert and text editor (ii) a leading planning GUI with built in UML modelling tools (iii) an object-based notation inspired by formal methods. We evaluate these three ap-proaches using process and product metrics. The results give insights into the strengths and weaknesses of the approaches, highlight lessons learned regarding knowledge encoding, and point to important lines of research for knowledge engineering for planning.

In addition, we discuss a range of state-of-the-art modelling tools to find the types of features that the knowledge engineering tools possess. These features have also been used for evaluating the methods used. We bench-mark our evaluation approach by comparing it with the method used in the previous International Competition for Knowledge Engineering for Plan-ning & Scheduling (ICKEPS). We conclude by providing a set of guide-lines for building future knowledge engineering tools.