Abstract: Bredeweg1992

[Bert Bredeweg - home page] | [Publications] | [GARP project - main page]

Publication details

B. Bredeweg. 1992. Expertise in Qualitative Prediction of Behaviour. Ph.D. thesis, University of Amsterdam, Amsterdam, The Netherlands.

Short abstract to each chapter

Chapter 0: Long Abstract

In this thesis we present a unified approach to qualitative prediction of behaviour, both as a knowledge level description and in terms of a detailed account of an implemented system that performs state of the art qualitative reasoning. The integrated framework is based on the three major approaches to qualitative reasoning: the component centred approach (deKleer, 1884), the process centred approach (Forbus, 1984), and the constraint centred approach (Kuipers, 1986).

Chapter 1: Introduction

Research on qualitative reasoning about the physical world has provided a number of major results, which at a first glance appear to be rather different (cf. Bobrow, 1984): the component centred approach (deKleer, 1984), the process centred approach (Forbus, 1984), and the constraint centred approach (Kuipers, 1986). Based on these initial approaches qualitative reasoning has gradually evolved into an independent area of research concerned with `automated reasoning about the physical world using qualitative representations' (cf. Weld, 1990).

Many researchers in the area of qualitative reasoning agree that there are similarities between the three main approaches, but little effort has been spent on uncovering what they are. The goal of the research presented in this thesis is to construct a theory of qualitative prediction of behaviour that encompasses the original approaches and that points out the essential conceptualisations of this problem solving task, thus enabling a better understanding of the similarities and differences between the original approaches.

Chapter 2: Approaches to Qualitative Reasoning

This chapter describes the state of the art of qualitative reasoning in three subsections. First an introduction to the field is given. The purpose of this introduction is to allow the reader to become familiar with the objectives in this area of artificial intelligence. The second section gives a detailed description of the three main approaches to qualitative reasoning. The last section discusses the main problems within the area of qualitative prediction. In this discussion we will concentrate on the problems concerning the three approaches mentioned before.

Chapter 3: Modelling Problem Solving

This section describes a theory for modelling problem solving behaviour, based on the $KADS$ methodology for building Knowledge Based Systems (KBS) (Wielinga, 1986; Breuker, 1987; Wielinga, 1988; Breuker, 1989; Hayward, 1987). We will use this methodology as a method for integration of the three main approaches to qualitative reasoning described in the previous section. It is therefore relevant that we give a detailed account of the important aspects of this methodology.

Chapter 4: Qualitative Prediction of Behaviour

This chapter describes an integrated conceptual framework for qualitative prediction of behaviour (cf. Bredeweg, 1988; Bredeweg, 1989; Bredeweg, 1990). In order to clarify the boundaries between prediction and other related tasks, such as modelling, the first section classifies the tasks relevant to behaviour prediction according to the criteria proposed by the theory discussed in the previous chapter. Having set these boundaries the description of the four layer model for qualitative prediction of behaviour is presented. In particular, the discussion of the framework focuses on an extended world view for representing partial behaviour models, the representation of parameter specific quantity spaces, an integrated set of parameter relations with additional functionality for causal value correspondence, and a transformation step between states of behaviour that encorporates an explicit selection and ordering of possible terminations.

Having described the problem solving roles (meta classes) in detail, the canonical inferences, from a conceptual point of view, turn out to be relatively straightforward. However, realising their problem solving potential in a computer program is a complex matter. The algorithms developed for that purpose are discussed in the next chapter which describes the design and implementation of the conceptual model.

Chapter 5: Problem Solving Behaviour in GARP

This chapter describes how the conceptual framework for qualitative prediction of behaviour can be transformed into a design model for the construction of a computer program. A design model in KADS consists of three views. The functional view (5.1) describes the functions that must be realised by the artifact. In particular, this section discusses some of the additional functionality that is required, besides problem solving, for implementation of the artifact (for example, the interaction with the user).

The most interesting problem to solve in this chapter is the construction of a behavioural view of the artifact. Algorithms must be developed that realise the problem solving behaviour required by the inferences in the conceptual model. Our solution to this problem is described in two sections. Section (5.2.1) describes how the meta-classes from the conceptual model are mapped onto design elements, and section (5.2.2) describes how these design elements are used by algorithms for realising the problem solving behaviour that is specified by the knowledge sources.

The physical view describes how the design elements and algorithms are composed into the different physical modules which constitute the actual artifact (5.3). The implemented program is called GARP, which is an acronym for General Architecture for Reasoning about Physics.

The last section describes two prediction models, one of the cooling mechanism of a refrigerator and one of heart diseases. Both models are implemented in GARP and illustrate important aspects of the problem solving behaviour manifested by GARP.

Chapter 6: Cognitive Plausibility

In this chapter we investigate the cognitive plausibility of the conceptual framework for qualitative prediction of behaviour (cf. Bredeweg, 1991). We compare think-aloud protocols of human subjects predicting the behaviour of a complex configuration of balances with a computer model of the same problem solving task, implemented in GARP. The contents of this chapter are as follows. The first section explains why cognitive plausibility is important. The second section discusses the nature of the problems on balances in more detail. The third section focuses on potential strategic knowledge that subjects may use during a behaviour prediction task. The fourth section describes the protocol analysis of the problem solving task and discusses to what extent the framework for qualitative prediction of behaviour fits the think-aloud protocol data. The last section summarises and further discusses the important results .

Chapter 7: Reflective Improvement

Strategic reasoning, in terms of the KADS four-layer model, is rarely employed in previous approaches to qualitative reasoning (see chapter 4). However, as discussed in the previous chapter, it is precisely this strategic layer that should contain the knowledge for reasoning about the knowledge represented in the other layers. This chapter describes how the strategic knowledge in the model of expertise can be operationalised as a reflective component `on top of' the artifact. In particular, this chapter discusses ways of improving the problem solving behaviour of GARP. These improvements are based on the strategic reasoning found in the protocols of human problem solving which was described in the previous section. In addition, the framework for strategic reasoning is based on the notion of reasoning about a problem solver.

Chapter 8: Conclusions and Outlook

In this thesis we have presented an integrated approach to qualitative prediction of behaviour, both as a knowledge level description and in terms of a detailed account of an implemented system that performs state of the art qualitative reasoning. Protocol analysis of human problem solving behaviour supports the cognitive plausibility hypothesis of the conceptual model and thereby its utility for knowledge acquisition. In addition, the approach has been augmented with an initial step towards reflective competence assessment and improvement. In the following sections the contributions of our research to these issues are further discussed. The last section points out a number of issues that are worthwhile for further investigation.

Last modified on May 9th, 2001

[TOP]