It is appropriate here to outline some of the work done in IMS on industrial applications of 'case-based reasoning', primarily using 'nearest neighbour matching' for precedent search, and induction as a means of learning from experience.

The science, or perhaps art, of 'case-based reasoning' has emerged from research work in what is sometimes called the 'artificial intelligence' research area of computer science. I have never been at ease with the latter label; those in the computer science domain who pursue 'AI' work I think are on a false trail. Despite this, some useful tools have come out of their work, and 'case-based reasoning' is one of them.

A 'case' is a structured record of a situation, some decisions made, and the outcome. An assembly of such 'cases' constitutes a 'case-base', and is in fact a record of experience. It can become a record of organisational experience, involving the experience of many people.

For such a case-base to be useful, however, the cases must share some common structure, such as to enable them to be described in the abstract by a manageable set of parametric values. Herein lies the art: delimiting the domain of the case-base so as to allow the spread of cases to describe adequately its variety and richness.

For example, a 'situation' might be defined in terms of features of a piece of land under consideration for forestry: slope, aspect, drainage, soil type, underlying geology, accessibility, altitude. The 'decisions' might relate to species or species-mix to plant, spacing, planting procedure, use of fertiliser. The 'outcome' is expressed in terms of value of successive thinnings and ultimate volume and quality of timber, at an assumed discount rate, over the period of the growth cycle.

Given a 'situation', one can search for 'similar' situations in the prior organisational experience. What decisions were taken, and what were the outcomes? The search is done by a 'nearest neighbour matching' procedure, which measures the 'distance' for the current situation from prior 'similar' decisions, in an n-dimensional non-euclidean space, using a 'similarity-matrix' approach to measuring 'distance', in dimensions where the values of the parameters are qualitative.

If the case-base has accumulated a large number of cases, one can analyse it inductively, using an 'induction engine', seeking to identify those cases which gave good outcomes, and use them to establish rules for decisions in various situations.

The 'AI' people go further and seek to set up 'adaptation' procedures, for making a past nearly-similar case fit a current situation. The philosophy behind this is to enable the system make its own decisions, without human intervention. In a business or professional decision environment, however, it is important that the human element be retained, becoming a skilled tool-user. The 'AI' aspect in this context needs to be abandoned, and the features of the 'case-based reasoning' tool understood and appreciated.

This combination of precedent-search and induction has proved to have many applications in engineering, and in maintenance of complex productive systems; it is showing signs of becoming useful in medicine.

Some navigational notes:

Copyright Dr Roy Johnston 1999