Rex Bing Hung Kwok

Translations of Ripple Down Rules into Logic Formalisms

The Ripple Down Rule (RDR) is a knowledge acquisition scheme which has been successfully used in large scale commercial applications. Most notable is the use of the scheme in building a knowledge base for interpreting pathology results. Unlike machine learning algorithms which construct theories for raw data, the commercial system using the RDR scheme requires some interaction with an expert to correct erroneous interpretations. However, the frequency of such interactions is very low. Furthermore, the simplicity of the interaction means that a knowledge engineer is not required when updating a knowledge base. A number of RDR variants have been devised for different applications. While Scheffer has given operational semantics to one variant of RDR, translations of RDR into classical logics has not yet been presented in the literature. This paper will show how two variants of RDR, single classification RDR (SCRDR) and multiple classification RDR (MCRDR), have a propositional or first-order core tied with case specific defaults. Such a translation will be used to highlight properties of the RDR scheme which make it successful. Similarities and differences in the two variants will be discussed and RDR revision will also be analysed.

Phylogeny, genealogy and the Linnaean hierarchy: a logical analysis.

Phylogenetic terms (monophyly, polyphyly, and paraphyly) were first defined in the context of a phylogenetic tree. However, reproduction is the background process that largely determines phylogeny. To establish a connection between genealogy and phylogeny, definitions of phylogenetic terms are presented and studied within a genealogical context. The correctness of the definitions is corroborated with results that show they satisfy the appropriate properties in the context of a phylogenetic tree. In an application of the definitions, a formal analysis shows why the monophyletic condition makes a Linnaean hierarchy entirely monotypic.

Using Ripple Down Rules for Actions and Planning

Much of the research into reasoning about actions is driven by a set of small benchmark problems. For such problems, specifying action behaviours within any particular logic of action is a simple task. In more complex domains this may not be the case. A simple and proven method for eliciting knowledge from experts — called Ripple Down Rules (RDR) — can be adapted for representing action descriptions and plans.

Creating Theoretical Terms for Non-deterministic Actions

Theoretical terms play a central role in many scientific theories and includes such terms as quark and lepton in physics. However, such terms do not refer to observables or the properties of observables. Due to their central role in many scientific theories, formalisations and implementations of scientific discovery should account for theoretical terms. Few methods have been developed within the field of Artificial Intelligence to account for such terms and little work on correctness has been done. This paper will define a formal method for creating theoretical terms based on observationally non-deterministic actions. Further, this paper will define a class of possible worlds models for which the method is provably correct.

The Coherence of Theories-Dependencies and Weights

One way to evaluate and compare rival but potentially incompatible theories that account for the same set of observations is coherence. In this paper we take the quantitative notion of theory coherence as proposed by Kwok et al. (Proceedings of the Fifth Pacific Rim Conference on Artificial Intelligence, pp. 553–564, 1998) and broaden its foundations. The generalisation will give a measure of the efficacy of a sub-theory as against single theory components. This also gives rise to notions of dependencies and couplings to account for how theory components interact with each other. Secondly we wish to capture the fact that not all components within a theory are of equal importance. To do this we assign weights to theory components. This framework is applied to game theory and the performance of a coherentist player is investigated within the iterated Prisoner’s Dilemma.

Coherence Measure Based on Average Use of Formulas

Coherence is a positive virtue of theories, and its importance in theory construction is beyond debate. Unfortunately, however, a rigorous and computable definition of coherence is conspicuously lacking in the literature. This paper attempts to remedy this situation by formalising this notion of coherence and suggesting a measure. Roughly speaking, we suggest that a theory is coherent to the degree that its members (read formulas) are required to account for the intended class of observations. This approach is motivated by Bonjour’s account of coherence. We also generalise this notion of coherence to work in the context of a potentially infinite long sequence of observations.

A Notion of Correctness with Theories Containing Theoretical Terms

This paper considers how theories containing theoretical terms can be tested with experiments which only give results in observable terms. An agent will be defined which makes predictions about the effect of actions. To test such an agent, the predictions of the agents will measured against an experimental frame. With the experimental frame a result will be presented demonstrating when some theoretical vocabulary is necessary. Further, the correctness of an agent will be defined and results showing the correspondence between beliefs containing theoretical terms and the experimental frame will be presented.