1. Identification Phase (scope of problem)
2. Conceptualization Phase (paper prototype)
3. Formulation Phase (paper prototype ® formal representation)
4. Implementation Phase (formal representation ® rewritten for tools)
5. Testing Phase – Check both "classic" test cases and "hard" boundary cases most likely problems.
1. I/O failures (VI)
2. Logic Rules (bad rules)
3. Control Strategy
6. Prototype Revision

Domain Expert

Knowledge engineer

Similar to prototyping

Change is healthy

1. Description (expert lectures or writes out task selection)
2. Observation
3. Introspection (knowledge engineer interviewing expert)

1. Expert may not have required knowledge of some areas
2. Expert may not be consciously aware of required knowledge
3. Expert may not be able to communicate knowledge to knowledge engineer
4. Knowledge engineer may not be able to structure knowledge for entry into knowledge base.

1. Knowledge is power
2. Knowledge is often inexact & incomplete
3. Amateurs experts slowly
4. Knowledge is often poorly specified
5. Expert systems need to be flexible
6. Expert systems need to be transparent
7. Separate inference engine and knowledge base level
8. Use uniform "fact" representation
9. Keep inference engine simple
10. Explicit redundancy

1. Recognized experts exist
2. expert do better than amateurs
3. Expert uses significant time
4. Cognitive tasks
5. Skill is routinely taught to neotypes (beginners)
6. Domain has high payoff
7. Task does not require common sense

R3 – decreasing reserve

R1 – Interest down

R5 –

R4 or R5

• what risk?
• interest up?

1. Assist in debugging
2. Inform use of current system
3. Increase confidence in advice
4. Clarification of terms and concepts used by system (help)
5. Increase users personal level of expertise (tutorial)

Car not starting ® Internal engine problem

why not? (counter factual reasoning)

what if? (hypothetical reasoning)

• Hand checking
• some formalism for examining relationship among rules
• and / or trees
• decision trees
• inference trees

M1

• 2500 rules
• Symbolic Variables
• Uncertain Knowledge
• Backward Chaining (mostly)
• Limits forward chaining
• Symbolic List processing
• Debugging Environment
• M1.exe, M1.cfg
• Explanation Facilities
• M1 seeks values by

1. Value of expression computed by M1
2. Value already in cache
3. Inference Rules
4. Prompt user for value

• Use lower case
• To comment /* */
• Readability is reduced if you use these commands. It is recommended that you do not use these.
• Prefix Eliminate dash on left
• Infix Dash on both sides
• Postfix Eliminate dash on right

multi-valued (preferred-color).
question(preferred-color) = ‘Do you prefer red or white wine?’.
legalvals(preferred-color = [red,white].

if x = y cf 30 then
   z = w cf 50.

initial data = [preferred-color, preferred-body, prefferred-sweetness].

automatic menu(preferred-color)

if best-body = light then
   recommended-body = light.

if best-color = white then 
   recommended-color = white.

if best-X = V then 
   recommended-X = V.

if recommended-color = red and
   recommended-body = medium and
   recommended-sweetness 
then wine = gamay.

if recommended-color = c and
   recommended-body = m and
   recommended-sweetness = s and
   wine (c,b,s = w )
then wine = w.

multivalued(wine).

multivalued(wine(color,body,sweetness)).

noautomaticquestion(wine(color,body,sweetness)).

if age N and N > 200 then
   category = adult.

if distance = D and 
   fair-per-mile = M and 
   A * D = R
then total-fair = R.

• Prescribe drug regimen for infectious virus and diseases

Design Parameters

1. Program is competent & easy to use
2. Handle a large, changing body of knowledge
3. Interactive
4. Must take time into account
5. Work with incomplete or uncertain information

Three Components

1. Consultation system(inference)
2. Explanation System
3. Rule acquisition /modification system

Premise

($and(same cntxt gram gram neg)
     (same cntst morph rod)
     (same cntxt hir aeroiste)
)

Consequent)

(conclude cntxt class enterobacteriacae tally .8)

Data Structures

context tree

• working memory

• AOV Triples
• (attribute object value)
• (ident orgi enterobacteriacae .8)

Control Structure

• backward chaining
• depth first

if there is an organism requiring therapy and consideration has been given to the possibility of other orgs requiring therapy.

then compile a list of possible therapies and select the best one from the list

Monitor- evaluate premise of current rule and discards it if false to try the next rule. {restrict by cntxt}

findout- gathers evidence for and against rule premise

1. Display rule being invoked (any point)
2. Record rule invocations and associate them with questions asked
3. Use rule index to retrieve particular rules to answer questions

1974- Panel of 5 experts approve 72% of 15 recommendations

1979 – Panel of 8 experts (5 medical examiners, 1 resident, 1 student, 1 research expert)
Compare MYCIN to actual, they looked at 10 cases, 52% of the time MYCIN was correct

if composition = (list of methods) and 
   error < 5 and
   stress > 0.5 and
   cycles > 1000
then SS stress = fatigue

1. Explain { rules}
2. Test {composition of current to stored results}
3. View {stored rules}

As you extend a rule base errors start to occur

1. Gaps in ruleset {combinations not covered}
2. Overlapping Rules {Inconsistent rules}
3. Rule obsolescence {new Discoveries invalidate rules}

1. Drawing false conclusions
2. Asking irrelevant questions

how/why

• maintain a long list of dependencies
• complicated data structure
• first attempt at knowledge acquisition

• Domain independent heuristics
• Add on to EMYCIN