Information Technology Reference
In-Depth Information
discriminated by human experts and tested by known approaches of deductive
reasoning and mathematical statistics.
Inductive learning can be classified into instance learning, observation and
discovery learning. The task of instance learning, also referred to as concept
acquisition, is determined by general concept description, which should explain
all given positive instances and exclude all given negative instances. These
positive and negative instances are provided by source of information. The
source of information is very extensive. It can be natural phenomenon or
experiment results. Instance learning which learns from classified examples
according to supervisor is a supervised learning algorithm.
Observation and discovery learning is also referred to as description
generation. This kind of learning will generate and explain the disciplines and
rules of all or most observations without the help of supervisor. It includes
concept clustering, construction classification, discovery theorems, expression
theories. Observation and discovery learning which learn from observations
without classification, or be discovered by functions itself, are unsupervised
learning algorithms.
Since inductive reasoning leads to complete knowledge status from definite
and incomplete knowledge status, it is a kind of non-monotonous reasoning.
However, inductive reasoning cannot verify whether the knowledge is right,
while non-monotonic logic provides theory foundation for us to handle
non-monotonic generative knowledge.
The basic idea of inductive principle is to formulate a scientific theory
through assumption on the basis of a great deal of observations. All observations
are singular proposition, while a theory is usually a universal proposition in
domain. There is not a logical inevitable implication relation between singular
proposition and universal proposition. They are usually default held for facts
cannot be observed. We use inductive assertion derived from inductive reasoning
as knowledge from the database. Furthermore, they are used as default
knowledge. When new proposition contract with them has been emerged, the
original default knowledge derived from inductive reasoning would be thrown
down so that the consistence of system knowledge can be kept.
A general definition of inductive learning from individual concept is as
following:
(1) Given an instance space constructed by all instances, each instance has
several attributes.
(2) Given a description language, the descriptive capability of the language
includes describing every instance (realized by describing its attributes) and
describing some instance sets, which is referred to as concept.
