Implications for Knowledge Engineering and Maintenance

The idea of introducing hidden and dynamic contexts into KBS research could potentially aid in understanding many of the historical and current problems in knowledge engineering and the maintenance of such systems. The fundamental issue, however, is how does a symbolic representation capture such information. Once a symbol is captured as a piece of knowledge, we not only are required to store a contextual relationship in parallel with that knowledge, we must also be able to alter it dynamically. Additionally, we should be able to find new contexts not expressed by an expert.

This thesis solves this problem by hybridising the symbolic representation with a function-fitting algorithm. The function-fitting algorithm learns the patterns of relationships between the partial symbolic representations of knowledge. Essentially working on Arbib’s idea that “…the representation of the world is the pattern of relationships between all its partial representations” (Arbib 1993, p 273).

2.4 Summary

This chapter briefly outlined two philosophies of knowledge. The first, physical symbol hypothesis and its later, less ad hoc, form of knowledge-level modelling

were discussed. The primary difficulty in these views was the conversion from contextually based to globally based knowledge and that these methodologies are inherently static in nature. Also discussed were a number of methodologies that use this interpretation of knowledge and how this presented further difficulties during knowledge acquisition and especially maintenance.

The emergence situated cognition (SC) view of knowledge addressed these issues, which was the second philosophy applied to AI discussed in this chapter. It was identified that knowledge representation methodologies need to be used that allow for dynamically changing knowledge. Ripple-Down Rules (chapter 3) is one methodology that has attempted to meet this SC view. Other methodologies that also use SC as their underlying philosophy, such as Personal Construct Psychology (PCP) and Formal Concept Analysis (FCA) were discussed.

These systems have performed exceptionally well in many areas where knowledge-level systems have repeatedly failed. Nevertheless, strong SC literature argues that such systems will not achieve true robustness and intelligence as context is too problematic to be represented symbolically. Finally, this thesis has suggested a clarification to what forms context can appear and suggest that for symbolic systems to succeed they must embrace all these contextual forms and that this meets the concerns of Strong SC. This redefinition of SC as it applies to AI, referred to as Intermediate SC is the primary philosophical influence behind the methodology in this thesis.

3

3

3

RRiippppllee--DDoowwnn

RRuulleess

If there’s a buzzing-noise, somebody’s making a buzzing-noise, and the only reason for making a buzzing-noise that I know of is because you’re a bee. … And the only reason for being a bee that I know of is making honey. … And the only reason for

making honey is so I can eat it. (Milne 1926, p 4)

Discussed in the previous chapter was a small selection of methodologies, both mainstream and context-based approaches to solving the knowledge acquisition bottleneck and maintenance issues prevalent in knowledge based systems. Generally, the mainstream approaches have become fixated on the development of knowledge-level models with a knowledge engineer. So much so, that they have lost sight of the observed and frequently reported fact that users want ownership of their knowledge (Freidson 1994; Ignizio 1991; Kidd and Sharpe 1987; Richards 1998a; 2000a). RDR, FCA (2.2.4.1) and PCP (2.2.4.2), on the other hand, represent a paradigm shift in the approach to KA and KM through the development of new context-sensitive representations for knowledge.

Compton et al. (1988) extended Popper’s (1963) theory of hypothetico deductive reasoning to the application of knowledge engineering. He suggested that experts do not provide information on their insight or how they reached a particular conclusion; but instead, they provide a justification for excluding the other possible hypotheses from within a particular context (Compton et al. 1988; 1989; Compton and Jansen 1988; Compton and Jansen 1990). It was suggested that these dichotomies between insight and justification (Compton et al. 1988; 1989) arise from a traditional misinterpretation of the form of knowledge provided by experts. This new hypothesis for knowledge engineering has resulted in the development and deployment of a new collection of methodologies and applications based on the knowledge acquisition and inferencing philosophy of Ripple-Down Rules (RDR).

After a discussion of the GARVAN-ES1 case study, which directly led to the development of RDR, this chapter will present a detailed description of the

basic RDR approach. Additionally, it will compare RDR with other methodologies and identify some of the recognised failings of RDR. Secondly, due to this thesis using Multiple Classification Ripple-Down Rules (MCRDR) as one of its base methodologies, it will discuss this primary adaptation of the basic philosophical ideas to the development of a multiple classification domain. These two KR structures are the basis for all RDR research and all the extension work presented in the final section of this chapter build on these fundamental methodologies.