Transcript print version

Dealing with Uncertainty
 The need to deal with uncertainty arose in “expert systems”

Code expertise into a computer system
Example:


Medical diagnosis: MYCIN
Equipment failure diagnosis in a factory
 Sample from MYCIN:

IF




The infection is primary-bacteremia AND
The site of the culture is one of the sterile sites AND
The suspected portal of entry is the gastrointestinal tract
THEN

There is suggestive evidence (70%) that the infection is bacteroid
 Expert systems often have long chains




IF X THEN Y … IF Y THEN Z … IF Z THEN W …
If uncertainty is not handled correctly, errors build up, wrong diagnosis
Also, there may be dependencies, e.g. X and Y depend on each other
Leads to more errors…
 Need a proper way to deal with uncertainty
How do Humans Deal with Uncertainty?
 Not very well…




Consider a form of cancer which afflicts 0.8% of people (rare)
A lab has a test to detect the cancer
The test has a 98% chance to give an accurate result
Mr. Bloggs goes for the test
 The result comes back positive
 i.e. the test says he has cancer
 What is the chance that he has the cancer?
 28%
 Afflicts experts too
 Studies have shown: human experts thinking of likelihoods do not reason
like mathematical probability
A
Metastatic cancer
No Link
Increased total
serum count
B
C
Coma
D
Brain Tumour
E
Severe
headaches
A
Metastatic cancer
… …
… …
… …
Increased total
serum count
B
… …
C
Brain Tumour … …
… …
… …
… …
… …
Coma
Serum
count
Brain
Coma
tumour
Yes
Yes
Yes
D
E
Severe
headaches
95%
Brain
tumour
headache
No
94%
Yes
70%
No
Yes
29%
No
1%
No
No
0.1%
Inference in Belief Networks
 Questions for a belief network:
 Diagnosis
 Work backwards from some evidence to a hypothesis
 Causality
 Work forwards from some hypothesis to likely evidence
 Test a hypothesis, find likely symptoms
 In general – mixed mode
 Give values for some evidence variables
 Ask about values of others
 No other approach handles all these modes
 Reasoning can take some time
 Need to be careful to design network
 Local structure: few connections
How Good are Belief Networks?
 Relieves you from coding all possible dependencies
 How many possibilities if full network?
 Tools are available
 Build network graphically
 System handles mathematical probabilities
 Case study:
 Pathfinder a medical expert system
 Assists pathologists with diagnosis of lymph-node diseases
 Pathfinder is a pun
 User enters initial findings
 Pathfinder lists possible diseases
 User can
 Enter more findings
 Ask pathfinder which findings would narrow possibilities
 Pathfinder refines the diagnosis
 Pathfinder version based on Belief Networks performs significantly better
than human pathologists