Whereas solutions to problems in these domains are difficult to program, learning may offer a
viable and effective alternative. In most inductive learning systems, human interaction is typically
limited to the gathering and compilation of examples (i.e., instances of the application to be learned).
As research on autonomous agents continues, this process itself may eventually become automated.
However, the strong knowledge principle [24], and early work on bias [16] suggest that examples
should be augmented by prior knowledge. Human learning, for example, is not the sole result of
exposure to random examples. Rather, built-in mechanisms (e.g., pain), and social structures (e.g., the
family, school) account for much of humans ability to efficiently learn complex problems. In many
cases, useful information is indeed available as, for example, an instantiation of domain knowledge or
commonsense. Though such knowledge is not necessarily correct, it can beneficially be used as a
learning bias to supplement inductive mechanisms (see, for example [7]). As a first attempt (and
currently the only viable one), this prior knowledge can be obtained from an expert or teacher. It may
be given to the system interactively or a priori, and can take many different forms (see, for example
[5]).