Information Technology Reference
In-Depth Information
could be terminated, or proceed to select a better assumption. Employing bias
approach, learning partial difference assumption will result in different inductive
leaps. Bias has two features(Utgoff, 1986):
(1) A strong bias is one that focuses the concept learning on relatively small
number of hypotheses. On the contrary, a weak bias is one that allows the
concept learner to consider relatively large number of hypotheses.
(2) A correct bias is one that allows the concept learner to select the target
concept. Conversely, an incorrect bias cannot select target concept.
Bias
Program designer
Hypothesis
Search Program
Knowledge
Trainer
Training Instances
Figure 7.1. Role of bias in inductive learning
Figure 7.1 shows the role of bias in inductive learning. From the figure we
can know that given any training examples with specific order, induction
becomes an independent variable function. When bias is strong and correct,
concept learning can select available objective concept. When bias is weak and
incorrect, concept learning is difficult because no guide can select hypothesis. In
order to transform a weaker bias, following algorithm could be employed:
(1) Recommend new concept descriptions to be added to the concept description
language through heuristics;
(2) Translate the recommendations into new concept descriptions formally
represented by concept description language;
(3) Assimilate newly formed concepts into the restricted space of hypotheses in a
manner that maintains the organization of the hypothesis space.
In the algorithm mentioned above, step 1 determines a better bias. Machine
executes transforming in step 2 and step 3, resulting in that new concept
description language is better than former description language.
To realize inductive learning, it is necessary to study a good bias. As for
fundamental problem of bias transform, it includes tasks about assimilate
