Furthermore, domain-specific systems that use instance-based learning algorithms have
performed extremely well in industrial applications. One example is ALFA (Jabbour,
Riveros, Landsbergen, & Meyer, 1987), a load forecasting assistant used by the Niagara
Mohawk Power Company of central New York State to estimate power load. ALFA uses
an instance-based learning algorithm to generate an initial prediction for power load, which
is then modified according to a rule-based domain theory. ALFA achieves the same accuracy
as experts but requires only two minutes to make load predictions (experts require two
hours). Another example is Clark's (1989) system for geologic prospect appraisal, which
is being used by Enterprise Oil to generate predictions of oil reservoir thickness and porosity
at a prospect site based on prospecting information gathered from nearby sites.
Using specific instances in supervised learning algorithms decreases the costs incurred
when updating concept descriptions, increases learning rates, allows for the representation
of probabilistic concept descriptions, and focuses theory-based reasoning in real-world applications.
However, no investigation has analyzed algorithms that use only specific instances
to solve incremental, supervised learning tasks.
In this article, we describe a methodology for instance-based learning (IBL) algorithms,
provide a geometric analysis to describe their generality and underlying intuition, address
two problems with this approach, and summarize convincing empirical evidence which
suggests that IBL algorithms perform well in applications to artificial and real-world domains.