Processing Certainty Factors
Two basic issues:
When to decide confidence is too low for rule to apply
How to compute certainty factors in rules
Certainty threshold
Within an inference engine, a threshold value typically is chosen for certainty factors
Values below threshold indicate conclusion is too uncertain to hold, so rule not used
Values above threshold indicate conclusion has acceptable certainty
Computing certainty factors
Rules state that a conclusion applies with given confidence, given hypotheses.
What confidence should be concluded if hypotheses are not completely certain?
Also, what confidence should be concluded if rule has multiple hypotheses, each of which is not fully certain?
Several approaches are possible to combine certainty factors.
Here is one approach, patterned after TMycin by Gordon Novak
Proceed in 2 steps
Determine confidence in rule conditions being met
Base confidence on rule conclusions on condition confidence and rule confidence factor
Confidence in Rule Conditions
If only one factor in condition, that gives the confidence
If several factors, connected with OR, take maximum confidence
If several factors, connected with AND, take minimum confidence
Confidence in conclusions, given hypotheses
Multiply rule confidence by hypothesis confidence
Notes:
Computed certainty factors can get low quickly, so one must be somewhat careful in choosing thresholds
A threshold confidence of 0.2 or so may be appropriate!
created May 13, 2003
last revised May 13, 2003
previous
next
