Information Technology Reference
In-Depth Information
requires that explanation must go on until it's calculable directly or the properties
can be observed directly. META-LEX considers the learning environment;
making cost of testing a property change with the system knowledge thereafter
operationality criterion becomes dynamic instead of static. PRODIGY further
improves the definition of operationality criterion.
9.8.1
Utility of PRODIGY
PRODIGY gets knowledge from experience analysis and interaction with experts.
Its learning aim is to obtain the goal concept with regard to problem solving such
as to learn goal concepts of successful, unsuccessful, and preferred. The typical
domain of problem solving is JSP (Job scheduling problem) which converts the
raw material by LATHE, CLAMP, POLISH operations as to become products.
The learning task for PRODIGY is to summarize under what circumstances a
particular operation is successful or unsuccessful. One peculiarity of this system
is that by adding some control rules, its efficiency is improved and the solution's
quality is amended. Meanwhile, new search which can not be obtained without
adding searching path considering efficiency can be found now. As for
operationality, PRODIGY requires that the learned knowledge should be able to
improve efficiency of problem solving and the control knowledge is not only
feasible but also applicable. Hence, the operationality includes applicability: the
learned control rule is not only feasible which means it can exert but also the rule
is capable of improving system performance. The applicability is defined as:
Utility=(AvrSavings × ApplicFreq) - AvrMatchCost
(9.1)
And
AvrMatchCost = average cost for matching the rule
AvrSavings = average saved time while applying the rule
ApplicFreq = frequency of application of the rule
When a control rule is learned, PRODIGY will preserve the statistical
information of its usage to decide its utility. If the utility value of the rule is
negative, this rule needs giving up. The operationality is to a large extent
represented by utility from this point of view. Therefore operationality can be
quantified and it should not be predefined but the value should be determined
dynamically during the system execution.
It is helpful for constructing general EBL and makes the evaluation of
operationality more mature to ascribe utility into operationality. Most of the
previous systems define operationality from the efficiency to recognize training
examples which is far more from adequate as for a general definition. For one
thing, the definition of operationality assumes concept description will be applied
to recognize training examples. Actually the concept description can be useful in
