Uncertain reasoning
Part of a sophisticated inference engines is a mechanism for dealing
with uncertain data and uncertain rules.
[Course
Home Page] [CIS
Dept] [Business School] [University
of Michigan]
Some issues:
Uncertainties do not obey the laws of probability.
Rules can be uncertain
- IF interest rates are dropping THEN stocks are a good investment
Data can be uncertain
- In the above, are rates really dropping? How much?
- This uncertainty can be based on humans(who hedge their judgments)
or noisy data.
There are different ways multiple uncertain rules combine to give a
single certainty:
- A rule may have several uncertain IF clauses (which need combining)
- There may be several rules, each of which makes the same conclusion
(with different levels of certainty)
- Uncertain data may combine with uncertain rules
There are various ways of combining these uncertainties:
- highest (optimistic?)
- lowest (pessimistic?)
- average (midground?)
- product (I don't see the motivation here)
- latest (good with thresholds)
- cumulative (accumulates corroborating evidence)
Value combo (set globally or for individual object | attributes) controls
how 2 or more distinct rules combine their uncertainties.
And-combo and Or-combo control how, within a single rule, the uncertainties
associatedwith different IF clauses are combined
Thresholds stipulate the minimum level of certainty for a rule to fire
(else it fires UNKNOWN)