KA-
Knowledge Acquisition
the most difficult tasks in developing expert system.
Different types of knowledge:
Types of Knowledge |
|
---|---|
Procedural |
Rules |
Declarative |
Concepts |
Meta-knowledge |
Knowledge about the other types of knowledge and how to use them |
Heuristic |
Rules of thumb |
Structural |
Rule sets |
Sources of knowledge:
expert - primary source of knowledge
end user - addition information
multiple experts - consult other experts
literature - reports, regulations, guidelines & books.
Knowledge Elicitation:
collect: -
acquire knowledge from expert
-
require effective interpersonal communication skills.
Interpret: -
review of collected information and identification of
key
pieces of knowledge
Analyze: -
forming the organization of knowledge and problem-
solving
strategies
Design: - design new techniques and collect new information.
Major difficulties in knowledge elicitation:
Experts may be unaware of the knowledge used.
Expert may be unable to verbalize the knowledge
Expert may provide irrelevant knowledge
Expert may provide incomplete knowledge
Expert may provide incorrect knowledge
Expert may provide inconsistent knowledge.
Roles of elicitation team members:
|
Elicitation Tasks |
---|---|
End-User |
Provide problem overview |
Expert |
Provide primary source of
knowledge |
Knowledge Engineer |
Collect the knowledge |
Interview with Experts
Guideline for obtaining initial cooperation
Remove Fear
convey that the expert is not a replacement, but will be an aid to help improve present performance.
Strip the mystery away and explain clearly what changes will occur with the introduction of the expert system.
Remove skepticism
provide a brief overview of expert system
provide a brief review of successful expert system projects on similar applications
provide frequent successful demonstration of the expert system
Establish Reasonable Goals
Don't oversell the system
Explain the capabilities as well as the limitations of the system.
Promote openness to change
explain how each team member's contributions can aid in the further development and acceptance of this technology.
Explain that only a temporary inconvenience will be experienced with the introduction of the expert system.
Provide Understanding of Expected Effort
describe the expected distributions of each member
convey to managers a description of each team member's project tasks
Convey Importance of Involvement
Make the team members aware that they are an important part of the project and without their effort the project has little chance to succeed.
Conducting the Interview
Follow the agenda:
should follow the agenda as close as possible (bring order to interview and offer best promises for achieving the session's objective)
an overly strict adherence to agenda can prevent unexpected discovery.
Deviation from agenda may create problem of not accomplishing the objective due to time limitation.
Type |
Purpose |
Form |
---|---|---|
Direct |
Obtain specific information on some known issue |
What does ... mean? |
Indirect |
Obtain general information on concepts and problem solving strategies. |
What issues are considered for
...? |
Probes |
Probe deeper into an established issue |
Can you explain ...? |
Prompts |
Direct interview into a new area |
Can you discuss ... ? |
Types of
Questions
Different types of questions commonly used in an
interview:
Guidelines |
Examples: |
---|---|
Avoid questions with more than one part to them |
Why are the pressure and the temperature of the pump ... |
Avoid leading questions |
Wouldn't you check the pump's temperature first? |
Do not phrase questions in the negative. |
Is it false that the water pressure is not high is not true? |
Avoid universal questions |
Do you always replace the pump when it is old? |
Avoid the use of terms foreign to the expert |
Do you also use forward reasoning to solve the problem? |
Watch for questions that use subjective words |
If the pump is really hot, do you shut it down fast? |
Guidelines for
forming a good question:
Use of funnel sequence techniques:
Active
Listening
e.g. paraphrase answers provided by experts to
clarify your understanding of the response.
Interruptions
interrupting expert during an answer for obtaining more information or redirecting the discussion.
General recommendation:
proceed with necessary interruption when expert is explaining declarative knowledge
avoid interrupt for procedural knowledge
Ending the interview
summarize key findings
review any outstanding questions raised during the interview and comment on how they will be answered.
Review any outstanding agenda items and discuss how they will be dealt with in the future.
Review future task for each team member before next meeting.
Maintain open communication lines for aiding future tasks.
Extend appreciation to members for their time and effort.
Unstructured Interview
allow expert to discuss a topic in a natural manner.
Get a conceptual understanding of the problem and insight into the general problem solving strategies used by the expert.
Usually used
in the early stage of knowledge acquisition.
Example:
The problem: problems with communication satellites causing loss in operation.
The solution: develop an expert system to monitor the performance of the satellite, detect any faults, and respond by either reconfiguring the satellite or by sending information to a ground station for request of corrective steps.
Dialogue:
KE: How do
you determine when the satellite is malfunctioning?
{starter
prompt}
DE: I notice that the message {CONCEPT} are garbled, or the BER {CONCEPT, domain vocabulary} is high {RULE}. This makes me sick when I think of all the money we invested in the thing and it still works worse than the radios I have at home {irrelevant}. And it always seems to come down to a couple of things that go wrong. The modulator {OBJECT} is the pits. This thing drifts off on us it seems every other day {HEURISTICS}. I think it mainly has something to do with its power supply {OBJECT}. Oh, wait a minute, that matrix switch {OBJECT} may even be worse {conflict}. It hangs up on us and sometimes doesn't make a good contact {HEURISTICS}. The output attenuator {OBJECT} can on a rare occasion also cause a problem {HEURISTIC}, and it's actually funny when it does. Ah ... I remember a time when ...
KE: Excuse me, can you tell me a little more why the matrix switch is such a problem? {prompt question}
Knowledge obtained:
Concepts: messages, BER
Objects: matrix
switch, output attenuator, modulator,
modulator power supply
Rules: IF Message
is garbled
OR BER is high
THEN A fault exists
Heuristics: modulator
drifts
matrix switch sometimes doesn't make good contact
output attenuator rarely a problem.
Guideline for an Unstructured Interview
Pose an initial prompt or indirect question "How do you ...?" or "What is ...?" This is called a starter prompt, which allows the expert to discuss a subject in general.
Interrupt the session with probe questions of the form, "Why is ...?" or "Can you please discuss ...?" to explore significant issues.
Be cautious with interruptions, particularly early in the project, when the expert may be providing valuable information that may be lost with the interruption.
Gain an
understanding of the important concepts, their relationships, and
general problem-solving methods from the transcript of the session
and use this information for planning future interviews.
Merits of Unstructured Interview
Advantage:
provides a general understanding of the problem.
Helps to identify important concepts and objects
Provide insight into general problem-solving methods
Allows for spontaneity in the interview which may give rise to previously unknown issues.
Disadvantage:
Information collected may be overwhelming, fragmented, or shallow.
Provides very little factual information.
Provides few details on concepts or objects
Provides few details on procedural information.
Structured Interview
maintain a focus on one issue at a time.
Elicits specific details on a given issue before moving on to other points.
Used later in the project when you have identified the problem's key topics from reports or earlier unstructured interview.
The problem: farmers seeking help from a regional agricultural extension service office.
The solution: An expert system was developed to provide readily available aid to farmers for growing and managing crops.
dialogue:
KE: In a prior session you mentioned that eliminating harmful pests is important. You also said that the first step in elimination is pest identification. Can you tell me what major characteristics you consider for identifying a pest? {focus prompt on characteristics}
DE: You can tell what kind of pest problem you have if you catch one of the little suckers and examine its appearance {CONCEPT}.Most farmers can identify the pest by looking at it, and ... ah ... or by inspecting the crop damage {CONCEPT}. Some of these guys will eat the leaves or roots {HEURISTIC} {RULE}. But before you try any pesticides you better be sure what it is. {HEURISTIC}
KE: Can you explain how you use the pest appearance in identifying the pest? {probe on appearance}
DE: You can look at the size {CONCEPT}, its color {CONCEPT}, or its shape {CONCEPT}. {RULE} Sometimes you can identify the pest from just one of these characteristics or other times you have to look at all of them. {HEURISTICS}
KE: Can you explain the size issue? {probe on size}
Knowledge obtained:
Strategies: View
the appearance of the pest first, then inspect
the crops for
damage.
Concepts: Pest
characteristics
appearance, size, color, shape
Crop
damage
leaf damage, root damage
Rules: IF The
size is something
AND The color is something
AND The
shape is something
THEN The pest is known
IF The leaf
damage is something
AND The root damage is
something
THEN The pest is known
Heuristics: Some
pests eat the leaves or roots
Before trying pesticides make
sure of the
identification of the pest
Sometimes pest
identification can be done using only
one pest characteristic.
Guideline for a Structured Interview
Choose a specific topic (focused prompt) to discuss.
From the discussion, identify new concepts.
Pick one of the new concepts (probe) and ask the expert to discuss it further.
Continue this process until enough details are obtained, then return to one of the new concepts mentioned earlier and ask for further discussion.
Gain an understanding of the important concepts and their relationships, rules and strategies from the transcript.
Merits of Structured Interview
Advantage:
Maintains a focus on a given issue.
Provides detailed information on the issue.
Provides insight into declarative knowledge used.
Study of one concept can lead to the definition of other unknown related concepts.
Provides structural relationships of concepts
Disadvantage:
Concepts unrelated to the interview's focus may not be found
You must be secure in your understanding of the important issues to direct the interview.
Provides only weak insight into procedural knowledge such as rules or problem-solving strategies.
Case Studies
Two primary ways:
Retrospective Case Study: the expert is asked to review the case and explain in retrospect how problem is solved.
Observational Case Study: ask expert to solve the problem discussed while KE observe.
Two case types:
Familiar cases: (typical cases) well known to the expert.
reveal typical knowledge used by the expert when solving the problem.
used early in the project to provide general insight.
Unfamiliar cases: study the problem in more detail, providing deeper problem-solving knowledge.
Retrospective Case Study Method
Select a case and ask the expert to explain how the problem was solved: e.g. Why was ... (the recommendation) ... given?
Expert should review the information of the case and try to justify how the problem was solved.
Example
The problem: Controlling of ventilation in coal mine.
The solution: An expert system was developed to aid coal mine managers in better controlling mine air ventilation system.
Case selection: A prior case was selected that recommended the installation of a larger main ventilation fan to improve the overall air flow throughout the mine.
Dialogue:
DE: I see that this mine had to go to a much large main fan {resuilt, OBJECT}. Let me see if they recently had a methane {OBJECT, also first one considered} buildup ... Um, no ... Now I see the problem. Look at some of these air flow {CONCEPT, also next one considered} numbers down there. Particularly at the face. These guys couldn't have been using any auxiliary fans {OBJECT} at the face. Let me see ... Yep, they have no extra fans there. This could be one of the major reasons they were told to get a larger main fan {HEURISTIC} ... it would probably have been cheaper too than buying some auxiliary equipment {HEURISTIC}.
Knowledge obtained:
Goals: Single recommendation 0 larger main fan
Strategies: Methane
level checked first, if no problem then check
in-mine air
flows particularly face air flows.
Concepts: fans, mine gases, faults, recommendations
Objects: main fans, auxiliary fans, methane
Rules: IF In-mine
air flow is low
THEN A fault exists
IF No face
auxiliary fans
AND Low face air flow
THEN Larger main
fan needed
Heuristics: Not
having auxiliary fans at the face could lead to a
need of a
larger main fan.
A larger main fan may be less costly than
purchasing
auxiliary equipment.
Guideline for Retrospective Case Study
Choose a typical case for typical information, or an usual one for more detailed information
Ask the expert to explain how he or she solved the problem
Record the expert's discussion for later study
During the expert's explanation, highlight important issues for further discussion following (or during) the session.
Be careful of interruptions. Interrupting problem-solving tasks can be disruptive.
Merits of Retrospective Case Study
Advantage:
Obtains information in the context of a working example
Provide problem-specific information
Does not interfere with the problem-solving activity.
Memory recall highlights the important issues.
Disadvantage
may provide incomplete information
provide few problem details
explanations can be inconsistent with actual past behaviour.
Observational Case Study Method
select problem and ask the expert for recommendation
ask the expert to "think aloud" while the problem is solved.
Familiar Case:
The problem: A transformer that is not routinely serviced or checked for incipient faults may cause severe damage.
The solution: An expert system was developed for incipient fault detection in transformers from the gas levels contained in a sample of transformer oil.
Case selection: The KE reviewed past cases, asking the technician who services it if it is a typical case.
Dialogue:
KE: I have
here a report of the test results on a transformer recently tested
by your technician. I would like you to tell me what
recommendations you would make regarding any problems with this
transformer. Also, as you study the report, please tell
me:
"What are your goals?"
"What issues
are important?"
"How are these issues
used?"
"What data do you use?"
DE: Well the first thing I want to determine if there is any problem with the transformer {GOAL}. There are a number of faults {CONCEPT} which can occur {check later}. Excessive buildup of certain gases {important CONCEPT} will usually tell me this... {RULE}. If it doesn't, I would look at ... {other CONCEPTS to explore}. Let me see... I see some nitrogen here, and a log of hydrogen {OBJECTS and FACTS}. Yea, I bet this baby has got an arcing problem {RULE and CONCEPT}. Let me check the past records on this transformer {STRATEGY}... This transformer has never shown this problem before {FACT}. Maybe its just a short term buildup which might go away if they adhere to good operating procedures {RULE}. I would recommend that they better control the load on this transformer {recommendation}, but I would like to see it retested in about ... let me think {explore later} 3 months {recommendation}.
Knowledge obtained:
Goals: Fault detected, recommendation given
Strategies: Gas levels checked first, if no problem then check ..., then past test data.
Concepts: gases, faults, recommendations, load control, retest period
Objects: nitrogen, hydrogen
Rules: IF Excessive
buildup of excessive gases
THEN A fault exists
IF Some
nitrogen exists
AND A lot of hydrogen exists
THEN Arcing
fault exists
IF It:s a
short term problem
AND Transformer correctly
operated
THEN retest in 3 months
Facts: Gas level data, retest time.
Guidelines
Choose a familiar case to study
Ask the expert to explain how he or she solved the problem with answers to questions such as:
what are your goal?
what issues are important?
how are these issues used?
what data do you use?
Ask the expert to "think aloud" while solving the problem
Record the concurrent protocol for later study
During the session, highlight important issues for further discussion following (or during) the session.
Merits of Familiar Case Study
Advantage:
Expert solves a real problem.
Obtains information in the context of a working example.
Common concepts, rules, and problem solving strategies emerge during the familiar case study.
Provide problem-specific information.
The "think-aloud" technique provides insight into the knowledge being used immediately rather than retrospectively.
Disadvantage:
Can miss important details unless introspection is used.
Explanations can be inconsistent with actual past behaviour.
Unfamiliar Case:
The problem: the transformer diagnostic again.
Case selection: The KE wanted to probe deeper into the problem. He wanted to uncover issues that may have not surfaced earlier in the project. Try an usual case.
Dialogue:
KE: (Present the case, ask the same question as before)
DE: Lot of gas {CONCEPT} data here {data-driven STRATEGY}. The hydrogen {OBJECT} and methane {OBJECT} is way up... so is the acetylene {OBJECT}... well somewhat. In fact, the percentage of combustible gas [CONCEPT} is too high... this baby is ready to blow {RULE}. Let me look at some of the past data {STRATEGY} ... um ... looks clean. I'd expect to see some combustible gases here {HEURISTIC}... something is happening really fast with this transformer {HEURISTIC}. Let me look at this data again {present data}. The ratio {CONCEPT} of methane to hydrogen is too high ... let me look at some of these others.
Knowledge obtained:
Strategies: Gas levels checked first, then past test data.
Concepts: gases, combustible gases, gas ratios
Objects: hydrogen, methane, acetylene
Rules: IF Percentage
of combustible gas is high
THEN transformer is ready to blow
IF Present
data have high levels of
combustible gases
THEN Past data
should have high levels of
combustible gases
IF Present
data have high levels of
combustible gases
AND Past data
don't have high levels of
combustible gases
THEN transformer
is deteriorating fast.
Guidelines:
have a second expert choose an unfamiliar case.
Ask the expert to solve the problem and answer questions such as: "What are your goals?", "how are these issues used?", "what data do you use?"
During the session, highlight important issues for further discussion following (or during) the session.
Use interruptions for introspection only if absolutely necessary.
Record the concurrent protocol for later study.
Merits of Unfamiliar Case Study
Advantage:
Forces the expert to use more basic knowledge, which in part helps to avoid the knowledge compilation problem.
Obtains information in the context of a working example.
Provide problem-specific information
Uncovers new details of previously known issues and can discover new issues.
Provides better insight into problem solving strategies.
Disadvantage
It can be difficult to choose a good unfamiliar case.
Capability of Elicitation Technique for Obtaining Knowledge
Knowledge Type |
Interviewing |
Case Study |
||||
---|---|---|---|---|---|---|
Unstructured |
Structured |
Retrospective |
Observational |
|||
Familiar |
Unfamiliar |
Familiar |
Unfamiliar |
|||
Facts |
Poor |
Good |
Fair |
Average |
Good |
Excellent |
Concepts |
Excellent |
Excellent |
Average |
Average |
Good |
Good |
Objects |
Good |
Excellent |
Average |
Average |
Good |
Good |
Rules |
Fair |
Average |
Average |
Average |
Good |
Excellent |
Strategies |
Average |
Average |
Good |
Good |
Excellent |
Excellent |
Heuristics |
Fair |
Average |
Excellent |
Good |
Good |
Poor |
Structures |
Fair |
Excellent |
Average |
Average |
Average |
Average |
Capability of Elicitation Technique for Avoiding Problems
Problem Type
|
Interviewing |
Case Study |
||||
---|---|---|---|---|---|---|
Unstructured |
Structured |
Retrospective |
Observational |
|||
Familiar |
Unfamiliar |
Familiar |
Unfamiliar |
|||
Unaware of knowledge |
Poor |
Fair |
Average |
Average |
Good |
Excellent |
Unable to verbalize knowledge |
Fair |
Fair |
Average |
Average |
Average |
Good |
Irrelevant knowledge |
Poor |
Average |
Average |
Average |
Average |
Good |
Incomplete knowledge |
Poor |
Average |
Poor |
Average |
Average |
Excellent |
Incorrect knowledge |
Average |
Average |
Poor |
Average |
Average |
Excellent |
Inconsistent knowledge |
Average |
Average |
Poor |
Fair |
Fair |
Excellent |
General Observations in Comparing Elicitation
Case study are better than interviewing techniques for obtaining procedural knowledge.
Structured interviews are the best choice for uncovering concepts, objects, and relationships.
Unstructured interviews are best for obtaining insight into general concepts and problem solving methods.
The observational case study method is, in general, better than the retrospective method for obtaining problem details and avoiding knowledge elicitation difficulties.
Familiar case studies are best for obtaining common domain concepts and typical problem solving methods.
Unfamiliar case studies are best for uncovering basic problem principles.
Unstructured techniques are poor for avoiding elicitation problems
In general, case studies are better than interviewing techniques for avoiding elicitation difficulties.
Unfamiliar case studies are best for avoiding elicitation difficulties.
Knowledge Analysis
Producing transcript of the recorded session.
Interpreting the transcript
Identify the key pieces of knowledge, the "chunks"
Use handwritten notes taken during the session to aid in identifying the key pieces of knowledge.
label each piece of identified information with the type of knowledge it represents.
Identify any issues that need further clarificaiton.
Analyzing the transcript
record each new piece of information with other similar pieces of information already discovered.
reference
each new piece of information to its source.
Review the body of knowledge collected with the expert to confirm the knowledge structures.
Highlight those areas that need to be pursued and use them in designing the next knowledge elicitation session.
Graphical
displaying the natural relationship between concepts or objects,
e.g. using cognitive maps.
Graphical representation of system's rules 0 inference network or AND-OR graphs.
Flow charting
the sequence of problem solving steps, or draw the decision trees.