n machine learning, the term ias" was introduced by Mitchell (1980) to mean any basis
for choosing one generalization [hypothesis] over another, other than strict consistency with
the observed training instances." Examples of such biases include absolute biases and relative
biases. An absolute bias is an assumption by the learning algorithm that the target function
to be learned is denitely a member of some designated set of functions (such as the set
of linear discriminate functions or the set of boolean conjunctions). A relative bias is an
assumption that the function to learned is more likely to be from one set of functions than
from another. For example, the decision tree algorithms (e.g., C4.5, CART) consider small
trees before they consider larger ones. If these algorithms nd a small tree that can correctly
classify the training data, then a larger one is not considered. The eld of supervised
learning has been described (Shavlik, J. and Dietterich, T.G., 1990) as the study of biases|
their expressive power, their computational complexity, and their sample complexity (i.e.,
the number of examples required to produce accurate generalization).