Search

When input a model of a planning domain and a problem to be solved, generative planning systems endeavour to generate a plan that is a solution to the input problem. There is a trade-off between the generality and efficiency of these systems which means that general planners tend to be inefficient. Our research has focussed on improving the efficiency of general planners. A popular approach to tackling this "efficiency versus generality" problem has been to embed systems with on-line and post-event learning components that use current and previous planning experience to increase subsequent performance. However, in this thesis we argue for a move away from this approach towards what we call off-line compilation, where all processing is independent of on-line planning. In support of this we have created two novel complementary techniques for off-line compilation, which we call PRECEDE and PROPOSE. In the course of our work they have been precisely defined, implemented and evaluated, and all aspects of this work are described in the thesis. PRECEDE is a method for generating goal orders that are based on necessary interactions between goals. PROPOSE is a method for generating iterative macro operators for all sorts of objects in a modelled planning domain, ft also generates sequences of constants that can be used to unwind the iterative parts of the macros. The PRE-CEDE goal orders can be used to select goals to achieve next and the PROPOSE iterative macros and enumerated sequences can be used to select ground operators for goal achievement during plan generation. These methods are off-line in the sense that all processing is performed a priori direct from the model of a planning domain. The compilation methods were evaluated statically, empirically and in comparison with established methods of on-line and post-event learning. For the empirical testing of off-line compilation an example planner was selected and extended so that it could use domain models that had been compiled by PRECEDE and PROPOSE. This planner was then used in a series of empirical tests which revealed that the techniques, both individually and in combination, consistently improve the performance of the planner (a speed-up of between 2 and 5 times faster across a sample of planning domains). Also, the results of static and dynamic comparison with established methods show that off-line compilation with PRECEDE and PROPOSE compares favourably with existing post-event and on-line learning methods.

Crowds have often been used in interactive entertainment applications to help provide context to virtual environment or enhance the experience of the interactive entertainment application for the user. This is particularly true in an interactive narrative, where many interactive narratives forgo the inclusion of crowds due to the effort required in order to create them or at best create crowds with simplistic programmed behaviours. Therefore, an approach to crowd simulation that solves the problems with including crowds in interactive narratives would allow designers of these application to consider the inclusion of crowds. The aim of this thesis is to create a novel approach to crowd simulation that solves four key problems of including crowds in interactive narrative applications. The main contribution of this thesis is a novel approach entitled Mise-En-Scene Region. This approach allows the control of agents inside an interactive entertainment application's virtual environment. The control includes: regions of the virtual environment specified with data on the expected population; behaviours an agent can perform in that region and types of agents that populate the region. During the run-time of an interactive narrative simulation, an agent autonomously performs a behaviour when entering these regions. As the simulation continues, an agent may become more or less likely to enter a region based on the region's current population. This approach is implemented in two interactive narrative applications that show the potential within interactive narratives. Finally, an evaluation has been conducted that shows the benefits of the approach in terms of qualitative and quantitative aspects.

The creation of planning domain models for automated planning is challenging, especially when non-technical domain experts are required for the creation of content. This is particularly true for the creation of domain models for Interactive Storytelling systems and games. AI planning can be used for the task of narrative generation and this could be utilised in such systems. Therefore a tool supported approach to the creation of narrative planning models that alleviates the requirement of specific domain modelling expertise and automates parts of the process would make narrative generation a more accessible technology in this context. The aim of this thesis was to develop a semi-automated approach to the creation of narrative planning domain models that automates the process of domain modelling and is accessible to non-technical authors. The approach taken aimed to use narrative synopses as an input for which a planning domain model can be acquired from. The contribution of this thesis is a novel approach for the acquisition of planning domain models from narrative synopses. The presented approach extracts the required planning information that is described by an input synopsis and from this automatically constructs a planning domain model that is representative of this information. Automated methods have been developed for the extraction of planning information that utilise having an author “in the loop” and exploit the contextual information available. A method for the automated construction of a planning domain model has been presented that is capable of reproducing the original input. This acquired planning model can then be generalised by an author using the default narrative control mechanisms that the model provides to produce a model capable of generating new story variants. The approach was implemented in a prototype system and evaluated to demonstrate the effectiveness of the approach.

This research project was initiated by the realisation that although the development of Computing learners' reflective skills is essential, reflection is a vague and a poorly defined concept. The dearth of literature (compared to other disciplines) with regards to supporting and defining reflection in the Computing discipline creates facilitation and assessment issues. This study starts with an investigation on the topic of reflection from disciplines who have a more mature practice of using reflection with their learners. The lessons teased out demonstrate the strong links between reflection and learning but also its key dimensions and complexity. The thesis progresses with the establishment of Computing teachers' views and perceptions of using reflection as a learners' evaluation tool. This investigation enabled the distillation of explicit variables (themes) considered key for the support and development of reflective skills in Computing. This work led to the introduction of the new concept of reflective development which entails one's own transformation and growth through a profound and inner meaningful change. Additionally, the thesis supplements existing literature on reflection by proposing a new reflective development framework to support teachers through the nurturing of their learners' reflective skills. Finally, the thesis explores how reflective development can be defined in Computing by analysing sets of learners' reflections and identifying development patterns and concepts which constitute reflective processes. This part of the study enabled the formulation of the reflective development model and attributes of good reflections in Computing. Action research was used throughout the study as a practice-based, problem-solving methodology as the research focuses on the enhancement of teachers and learners' practice but also contributes to educational theories. Four action research cycles were required to formulate the main contribution to knowledge of the thesis i.e. the reflective development concept including its framework and model.