Information Technology Reference
In-Depth Information
generating knowledge and question acquisition, computing initial bias,
approaches of goal freedom and goal sensitivity, etc. It needs to be studied
further.
7.4 Version Space
Version space takes the whole rule space as initial assumed rule set
. According
to information of training examples, it makes generalization or specialization to
set, increasingly reducing the set
H
is converged into rules that only
include quest. The term version space is used to refer to this set because it
contains all plausible versions of the emerging concept.
In 1977 Mitchell pointed out that rules in rule space can build partial order
according to their general degree. Figure 7.2 shows a partial order in a rule space,
where TRUE means that there is no condition, and this is the most general
concept. Concept
∃
x
: CLUBS(
x
) means that at least a club is more specific than
the former. Concept
∃
x,y
:CLUBS(
x
)
∧
HEARTS(
y
) means that at least a club
and a heart exist and they are more specific than the former. Arrows in the figure
point to more general concept from specific concept.
H
. Finally
H
Figure 7.2. A partial order in rule space
Figure 7.3 is the sketch map after ordering the general rule space. The highest
point in the figure is the most general rule (concept). It is a point without
description, i.e. a point without condition. All examples are in accordance with
the concept. Points on the lowest line are correspond concepts of positive training
examples. Each point corresponds to a positive example. For example, every
example shows suit and rank of a card C, such as:
SUIT(C, clubs)
∧
RANK(C,7)
This is a positive training example, at the same time it is the most specific
concept. Concept RANK(C,7) are points at the middle of rule space. It is more
specific than no description, and more general than positive training example.
