Algorithms that employ these representations were improved by either storing and using
specific instances or by supporting partial matching techniques. For example, Utgoff (1989)
showed how the updating costs of incremental decision tree algorithms can be significantly
decreased by saving specific instances. Similarly, Volper and Hampson (1987) showed that
the perceptron's learning rate can be significantly increased in some applications by using
specific instance detectors. Finally, Michalski et al. (1986) extended AQI1 (Michalski &
Larson, 1978) to employ partial matching strategies to support the description ofprobabilistic
concepts as defined in Smith and Medin (1981). The topic of this article is instance-based
learning algorithms, which use specific instances rather than pre-compiled abstractions
during prediction tasks. These algorithms can also describe probabilistic concepts because
they use similarity functions to yield graded matches between instances.