Robotics Reference
In-Depth Information
their thought processes because much of their thinking is almost sub-
conscious, or it might appear to them to be trivial or obvious. One of
the experts being debriefed for a well-known expert system project called
DENDRAL, was shown the data relating to a particular chemical com-
pound and observed: “It's a ketone”. 48 The knowledge engineer asked:
“How do you know it's a ketone?” and was somewhat puzzled by the
reply: “Well—look at it. It's gotta be a ketone.”
Before he can extract knowledge from the experts, a knowledge en-
gineer should first become at least somewhat familiar with the domain
of interest, possibly by reading an introductory text and/or by talking to
the experts. After this induction it is often useful to set each expert sev-
eral example problems, asking them to explain aloud their reasoning as
they solve each problem and to mention and explain any rules of thumb
that they employ. The knowledge engineer can often extract some gen-
eral rules from these explanations and then check them with the experts
before they are encoded for the system. The experts are further observed
and debriefed while they are engaged on many more relevant tasks. Again
the experts are asked to verbalize their thought processes as they work on
the tasks, allowing transcripts of their verbalizations to be coded by the
knowledge engineer into an appropriate format for the knowledge-base.
Realistically this whole process might take several person-weeks, months
or even years, in order to extract and refine sufficient rules to enable the
creation of a powerful expert system.
Most experts typically cannot afford to devote several weeks or
months to a research undertaking of this type, and they don't want to;
they prefer to work on their research in their own chosen field, so as the
science of expert systems has developed, computer-based tools for knowl-
edge acquisition have been created. Some of these tools work directly
with human experts to elicit knowledge and structure it appropriately to
operate within an expert system. But the elicitation of expert knowledge
and its effective transfer to a useful knowledge-based system is complex
and involves several diverse activities. Therefore, starting in the 1990s,
fully-automated methods of knowledge discovery and acquisition have
been developed.
One automated approach is to employ neural networks to induce
rules by generalizing from existing examples of their use. From a set of
observations,
rules are derived automatically,
rules that relate the
48 A particular type of chemical compound.
Search WWH ::




Custom Search