Information Technology Reference
In-Depth Information
All the leaf nodes in the explanation structure is operable, a case in point is
Made-of(
x,m
) and Spec-Grav(
m,d
). Some intermediate nodes might be operable,
in the above figure, Density(
) is operable. Therefore the operable boundary
can be acquired in the figure. All the nodes below the dashed line are operable.
As for the intermediate operable nodes, sub-explanation can be deleted from the
explanation structure. In the above figure, Made-of(
x,d
) and Spec-Grav(m,d)
should be deleted. This is called “Prune of operationality”. The following rules
are acquired after explanation in Figure 9.13:
Volume(
x,m
x,v
)
∧
Density(
x,d
)
∧
X(
v,d,w
)
∧
Isa(
y
,rect-solid)
∧
Made-of(
y
,lucite)
∧
(
)
As for the disposal of operationality criterion, it is a little too complex. Only
the confirmation of operationality criterion involves more aspects and this will
affect the performance of the learning system directly, consequently it is a rather
important problem.
There is no precise definition of operationality criterion among lots of
systems. Many systems take it for granted that operationality criterion should be
independent and static. The static standard proposed by DeJong & Mooney can
not satisfy the requirements and the operationality criterion should be the
function that can prove structure in terms of an example, that is to say,
operationality criterion should be dynamic. For instance, predicate of end-point
for the generalized explanation structure should be operable; the operationality
can also be defined by a theorem prover mechanism. EBG can even be applied to
decide operationality. Anyway, the operationality criterion must reflect the
computation cost and benefit of learning, and make sure the computation cost
ought not to higher than the benefit. However, the initial EBL leaves out this. It is
generally believed that knowledge learned from EBL problem solver system can
certainly improve performance of the system, but this is too optimistic and
simply. Generally speaking, EBL often implements heuristic strategy, which can
not guarantee the improvement under all circumstances. For instance, the
learning of STRIPS MACROPS is actually deteriorates efficiency. If the
accumulated time of testing premise of macro-operator exceeds the saving time
by using macro-operator for searching, then the overall performance is lowered.
Hence it is obvious that operationality is no doubt completely independent with
system.
There is no doubt EBG proposed by Mitchell et.al is an intuitionist way to
test the practically of explanation explicitly by operationality criterion. But as
DeJong and Mooney pointed out the above operationality criterion only include a
small portion of predicates and only the new knowledge can be assessed directly
but not test its usefulness. As in the example of EBG, operationality criterion
w
,10)
ŗ
Safe-to-Stack(
x,y
