Title: COSC 4350 Artificial Intelligence
1COSC 4350Artificial Intelligence
- Uncertain Reasoning Basics of Probability Theory
(Part 1) - Dr. Lappoon R. Tang
2Overview
- What is uncertainty reasoning?
- Why is it an important topic in AI?
- Brief history of uncertainty reasoning
- Basics of probability theory
3Readings
- R N Chapter 13
- Sec 13.1 (Focus on the motivation for uncertainty
reasoning) - Sec 13.2
4What is uncertainty reasoning?
- Performing inference under uncertainty (i.e.
incompleteness of information) - Making rational decision that maximizes expected
gain even though not all information necessary
for making a perfect decision is available - Example Stock markets, gambling,
- Need a way to characterize uncertainty (e.g.
degree of uncertainty, likelihood, etc) - Probability theory is our best choice of formalism
5Why uncertainty reasoning?
- There are situations in real life that require
reasoning and decision making based on uncertain
evidence and incomplete information - Classical logic only allows conclusions to be
strictly true (or false) - Problems
- Does not account for uncertainty
- Cannot weigh and combine evidences
6Logical vs Uncertain Inference
- Logical
- Necessary
- Exact
- Logical formula
- (including True / False)
- Theorem proving
- Uncertain
- Not required
- Estimation
- Probability
- Probability calculations
- Features
- Complete Information
- Quality of conclusion
- Representation of conclusion
- Inference Mechanism
7History of uncertain reasoning
- MYCIN (1975)
- One of the first expert systems ever created
(applied on diagnosing bacterial infections),
introduced rules with certainty factors. Uses
some pretty ad hoc ways of combining evidence. - Dempster-Shafer Theory (1976)
- Associates upper and lower bounds on beliefs and
has clearer foundations in probability theory - Bayesian Networks (1986)
- Use directed graphs to explicitly represent
causal dependencies and use inference processes
based directly on Bayesian probability theory - Most popular form of probabilistic reasoning
nowadays - Combining logical and probabilistic reasoning
(mid or late 80s - present) - Explores ways to combining the strengths of the
two paradigms into one, they come in various
flavors - Probabilistic relational models Friedman, et.
al, 99 - Probabilistic logic programming Ng and
Subrahmanian, 92 - A lot more
8Probability Theory Basic Concepts
- True Story
- Many years ago, a student at Stanford
- University wrote the San Francisco Weather
- Bureau to ask what it meant by a 60
- chance of rain, did it mean it would rain
- over 60 of the area? Did it mean it would
- rain 60 of the time?.
9Probability Theory Basic Concepts
- The Weather Bureau replied that it meant
- that there were ten men at the Weather
- Bureau and six of them thought that it was
- going to rain
10Probability Theory Basic Concepts (contd)
- The lesson We need a way of defining the chance
of seeing a certain outcome - The sample space is the set of all possible
outcomes (in an experiment) - Example in tossing two coins, the set of all
possible outcomes (H,H), (H,T), (T,H), (T,T) - An event is a subset of the sample space
- Example in tossing two coins, the event at
least one head appears (H,H), (H,T), (T,H) - The sample space itself is also an event!
- An event is a particular collection of outcomes
11Probability Theory Basic Concepts (contd)
- Let S be the set of all possible outcomes (i.e.
the sample space). The unconditional or prior
probability that an event A occurs P(A) (i.e. one
of the outcomes of A happens) is given by - Where A is the number of outcomes in A and S
is the sample space
12Probability Theory Basic Concepts (Examples)
- Example the set of all possible outcomes in
tossing two fair coins (H,H), (H,T), (T,H),
(T,T), the event A two coins turn up the same
side consists of the set of outcomes (H,H),
(T,T) - So, p(A) A / S 2 / 4 0.5
13Probability Theory Basic Concepts (Examples)
- Example the set of all possible outcomes in
tossing two fair dices (1,1), (1,2), ,
(6,6), the event A two dices sum to six
consists of the set of outcomes (1,5), (2,4),
(3,3), (4,2), (5,1) (there is no (0,6), so, it
is not an outcome ?) - So, p(A) A / S 5 / 36 0.139
14Probability Theory Basic Concepts (Basic Laws)
- If S is a sample space, and A is the complement
of A (i.e. A S A), then - P(S) 1 (one of the outcomes of all possible
outcomes has to occur) - P(A) 1 P(A) (an outcome either occurs in an
event or it does not in which case it occurs in
the complement of the event) - P( ) 0 (you cannot have no outcome)
- For any event A, 0 lt P(A) lt 1 (0 means event A
can never happen, 1 means A will always happen) - If A is a subset of B, then P(A) lt P(B) (the
chance that an outcome occurs in an event with
more outcomes is higher)