On the efficiency of meta-level inference

Abstract

In this thesis we will be concerned with a particular type of architecture for reasoning systems, known as meta-level architectures. After presenting the arguments for such architectures (chapter 1), we discuss a number of systems in the literature that provide an explicit meta-level architecture (chapter 2), and these systems are compared on the basis of a number of distinguishing characteristics. This leads to a classification of meta-level architectures (chapter 3). Within this classification we compare the different types of architectures, and argue that one of these types, called bilingual meta-level inference systems, has a number of advantages over the other types. We study the general structure of bilingual meta-level inference architectures (chapter 4), and we discuss the details of a system that we implemented which has this architecture (chapter 5). One of the problems that this type of system suffers from is the overhead that is incurred by the meta-level effort. We give a theoretical model of this problem, and we perform measurements which show that this problem is indeed a significant one (chapter 6). Chapter 7 discusses partial evaluation, the main technique available in the literature to reduce the meta-level overhead. This technique, although useful, suffers from a number of serious problems. We propose two further techniques, partial reflection and many-sorted logic (chapters 8 and 9), which can be used to reduce the problem of meta-level overhead without suffering from these problems.