Information Technology Reference
In-Depth Information
which the explanation structure holds, stated in terms that satisfy the
operationality criterion. This is accomplished by regressing the goal concept
through the explanation structure. Actually, the first step separates the relevant
properties and irrelevant properties of instance apart, while the second step is to
analysis the explanation structure.
(1) Generate explanation. After an instance is given, system begins to solve the
problem. If the inference is executed backwardly from goal, relevant rules
that match the consequents need to be searched from domain knowledge base.
When these rules are found, goals are set to be consequence and rules become
premise, and then the relationship is recorded. Premises of these rules are
considered to be sub-goals, the inference goes on until the problem solving
completes. Once the solution is acquired, the goal is proofed that it can be
satisfied and the explanation structure of cause and effect is obtained.
There are generally two ways to construct explanation structure. The first one
is to make a set of operators for every inference to construct action sequences to
form explanation structure; the other one is traversing the structure of proof tree
in top-down way. The former is general, leaving out description of some facts
while the latter is more specific, every fact appears in the proof tree. The
explanation can be constructed in the process of problem resolving and it can also
be accomplished after the problem solving. These two methods are so called
learn while study and learn after study.
(2) Generalize core event of the explanation structure. In this step, generally
adopted way is to convert constants into variables, i.e. to change the concrete
data into variable for these examples and omit some unimportant information;
only preserve key information for solution. Production rules are created to
obtain control knowledge for generalization.
To see how the EBG method works more concretely, considering the problem
of learning the concept SAFE-TO-STACK
(x, y
). The task is to learn to recognize
pairs of objects <
> such that it is safe to stack x on top of y. The instance of
SAFE-TO-STACK is as follows:
x, y
Given
:
• Goal Concept: Pairs of objects <
x, y
> such that SAFE-TO-STACK (
x, y
),
where SAFE-TO-STACK(
x,y
)
⇔
NOT(FRAGILE(
y
))
∨
LIGHTER(
x,y
).
• Training Example:
ON (OBJ1, OBJ2)
ISA (OBJ1, BOX)
ISA (OBJ2, ENDTABLE)
COLOR (OBJ1, RED)
COLOR (OBJ2, BLUE)
