Learning forms an essential part of artificial intelligence applicatio
ns. Databases of example ''cases'' are essential to the development an
d assessment of learning systems. insufficient examples make it diffic
ult or impossible to compare variations of the learning method. Suffic
ient examples are rarely available to assess a learning system thoroug
hly. Simulation represents a means of producing data based upon a defi
ned system. In this paper, a simulation for assessing the characterist
ics of learning systems is described. The simulation aims to generate
data as an actual knowledge-based system (KBS) observes and stores dat
a. A notional model is developed to mirror what is known of part of th
e actual target domain of a particular KBS. Significant results materi
alizing from simulated data include a quantitative comparison of learn
ing and testing on the same and disjoint data sets. Simulated data is
used to show that the use of the same data for learning and testing fr
equently reduces diagnostic accuracy when learnt knowledge is applied
to new data. Copyright (C) 1996 Elsevier Science Ltd