CSRL

1
CSRL

• CSRL

• CSRL

Why develop yet another language?
2
Because it looks like this
(specialist gas (declarations (superspecialist
fuel)) (knowledge-groups (summary (table
(conditions (ask-tfu? (did the problem start
after the last fillup?)) (ask-tfu? (has the
problem gotten worse since the last fillup?))
(ask-tfu? (have you tried a higher grade of
gas?))) (match (if (t ? ?) then 3)
(if (f t ?) then 3) (if (f f f) then 3)
(if (u u u) then 0) (if (? ? ?) then
-3))))) (messages-received (establish
summary) (refine usual-refine)
(establish-refine usual-establish-refine)))
3
Why is the code like that?

• Programming language can model the way experts
  reason.
• Two ways medical experts diagnose a case
• 1) Can diagnose easy cases
• 2) Or they know enough to send the patient to the
  right specialist.

4
The Plan

• What is CSRL exactly.
• Issues like uncertainty, handling multiple
  diseases, generic tasks.
• Control System.
• Real uses of the system
• Compare to Mycin and Internist
• of slides 32

5
Quick History

• MDX was developed first
• Diagnoses Cholestasis (a liver disease)
• Automech was a system developed after using CSRL.
  
• Diagnoses fuel problems
• Main difference from other systems is

6
Classification Hierarchy of Automech

More Specific
7
Terms

• Specialist each node in the classification
  tree.
• Hypothesis a possible disease
• Establish a hypothesis is confirmed
• Knowledge Group where the knowledge is stored
  in the form of symbolic certainty factor

8
Declaring a specialist

• (Specialist BadFuel
• (declare (superspecialist FuelSystem)
• (subspecialists LowOctane WaterInFuel
  DirtInFuel))
• (kgs)
• (messages)
• )

9
Specialists Can..

• Activate subspecialists to confirm a hypothesis
  (Establish Refine)
• Confirm a hypothesis (kgs)
• Rule out a hypothesis (kgs)
• Direct the reasoning of other specialists
  (Suggestion messages)

10
Messages (Establish Refine)

• (Establish (if (GE relevant 0)
• then (SetCertainty self summary)
• else (SetCertainty self relevant)
• (Refine (for specialist in subspecialists
• do (Call specialist with Establish)
• (if (? Specialist)
• then (Call specialist with refine)))

11
Message Establish
Fuel System Problems
(Establish Mixture)
(Rejected Mixture 3)
Fuel Mixture Problems

12
Messages Refine
Fuel System Problems
(Refine Vacuum)
Fuel Mixture Problems
Vacuum
(Establish Carburetor-gasket )
(Establish Vacuum-hoses )

Carburetor-gasket
Vacuum-hoses
13
Messages (Suggestion)

• Suggest what the subspecialist should consider
• Done for efficiency

14
Message Suggestion
Fuel System Problems
(Refine Vacuum (suggest Carburetor-gasket))
Fuel Mixture Problems
Vacuum
(Establish Carburetor-gasket )

Carburetor-gasket
Vacuum-hoses
15
Knowledge Groups

• Match data against patterns to produce certainty
  factor.
• Certainty factor is a symbolic value between 3
  and 3. (more on this later)
• 2 or 3 mean Established

16
Final Outcome
Fuel System Problems
((Established Carburetor-gasket 3) (Unknown
Vacuum-hoses 1 ))
Fuel Mixture Problems
Vacuum
(Established Carburetor-gasket 3)
(Unknown Vacuum-hoses 1 )

Carburetor-gasket
Vacuum-hoses
Question asked of user Can you see cracks in the
carburetor gasket? And Is the diagnosis finished?
17
Output

• Output
• (answer (rejected delivery 2)
• (rejected mixture 3)
• (established vacuum 3)
• (rejected bad-gas 3)
• (established carburetor-gasket 3)
• (unknown vacuum-hoses 1)
• (established fuel-problems 2))

18
Key things

• Knowledge is used in the right contexts
• Knowledge is directly involved in the problem
  solving
• Suggestion messages could have made it faster

19

• CSRL

• CSRL

• Issues

20
Uncertainty

• No global calculus keeps added uncertainty error
  to a minimum
• Abstractions chosen by expert in context

Disease S1 S2 Certainty
D1 - High
D2 - Low
21
Control

• Focus of control changes based on which
  specialist is active
• Control is distributed
• A specialist may be activated by a parent
• By another specialist through suggestion rules

22
BlackBoard

• Why?
• Interface between knowledge sources
• Allow for concurrency
• Hold the state of the system
• Examples
• One disease may be secondary to another
• The disease may have more than one cause

23
BlackBoard

• 4 different sections of the blackboard
• Active-hypotheses
• Established-hypotheses
• Rejected-hypotheses
• Suspended-hypotheses hypotheses for which a
  specialist has not found sufficient evidence to
  justify pursuing them.

24
BlackBoard

• A specialist activates and is placed in the
  Active list.
• If it gets establishes, it will place its name
  in Established-Hypotheses with its
  manifestations and recommendations for its
  subspecialist.
• It will activate its successors.
• Remove itself from Active-List
• If it rejects, it will enter its name on the
  Rejected-Hypotheses list along with all the
  negative evidence. Then deactivate.
• If a specialist suspends itself, it places itself
  on the Suspended-Hypotheses followed by a list of
  manifestations.
• Note If all the subspecialists of a specialist
  have rejected or suspended themselves then the
  parent must also be rejected or suspended.

25
BlackBoard

Active Hypothesis (none)
Established-hypotheses Liver(A1,A2,A3) Lung(S1,S2,S3) BrokenLiver(G1,G2)
Rejected-hypotheses Heart(H1, H2,H3)
Suspended-hypotheses Tumor (G1,G2,G3) Stone(S1,S2,S3)
26
BlackBoard

Active Hypothesis (none)
Established-hypotheses Liver(A1,A2,A3)
Rejected-hypotheses Heart(H1, H2,H3)
Suspended-hypotheses Tumor (G1) Stone(A1,H2,A3)
To be explained (G1)
27
Overview Critic

• Assembles a composite hypothesis that best
  explains the data.
• Ensure that all suspended specialists
  manifestations have been explained by established
  hypotheses
• Or
• How? Take list of unexplained manifestations(sympt
  oms) and see if suspended specialists can explain
  them.

28
Multiple Diseases

• Problem
• Diseases with Additive Symptoms
• Cardiac cirrhosis causes elevation of bilirubin
  up to 3mg/dl
• Cholestasis can account for excess of 5mg/dl
• What if the blood test shows 9mg/dl?

29
Multiple Diseases

• Solution Add more message types, can-account-for
  and re-establish
• (Established Cholestasis 2) likely
• (Established (CardiacCirrhosis 1) possible
• Then, higher level specialist integrate results.

30
Multiple Diseases

• Cholestasis returns
• (establishedAt 2
• (canAccountFor (bilirubin 9))
• Liver responds
• (re-establish
• (assume (bilirubin 6))

31
Generic Tasks

• CSRL is classification and hypothesis matching
• Not heuristic classification
• Because does not do data-abstraction
• (ex. Blood count 5 Blood count low)

32
Strengths and Weaknesses

• Strengths
• No global calculus
• Readable and code that models experts reasoning.
• Weaknesses
• Limited to classification and hypothesis
  matching.
• Not a full programming language.

33
Other Systems Automech

• Domain Car fuel system diagnosis
• Organization 34 specialists.
• Results
• Easy for non-computer experts to understand
• Easy to debug
• Took 4 people five man-months of time to develop,
  30 was expert time.

34
Other Systems Red

• Red anti-body identification
• Organization Three components.
• CSRL is used to rule out antibodies

Overview
DB
CSRL
35
Other Expert Systems

• Mycin - uses a hierarchy, but limited to
  knowledge that can be expressed as rules
• Internist decision based heavily on scoring
  function for manifestations.
• Casnet performs heavy calculations. Uses a
  causal network to map states to tests.

36
Conclusions

• CSRL organizes knowledge in a hierarchy of
  specialists
• Knowledge is directly involved in the problem
  solving
• It is flexible to handle issues such as multiple
  diseases.

37
Discussion

• CSRL

• CSRL