Retrieval Issues in Real-World CBR Applications How far can we go with discrimination-nets?

Abstract

In this paper we present a proposition and ponder a question. We propose that a useful perspective on analogical reasoning and CBR is to consider them on a continuum of abstraction of remindings. This is an alternative to the conventional view where CBR and analogical reasoning are seen as separate endeavours with analogical reasoning dealing with remindings between domains and CBR concerned with remindings within one domain. The question is how far towards the abstract end of the continuum can the index-based retrieval techniques that are effective in CBR be used (eg. discrimination networks). We are considering episode retrieval as a two stage process; the first stage being the initial filtering of the case base, and the second stage selecting the best case from this candidate set. We focus on the base filtering stage and conclude that discrimination networks are adequate for quite complex applications. However, problems arise when the system is required to support remindings at different levels of abstraction.