Uncertain Reasoning

Why Uncertain Reasoning?

Model Of Intelligent Agent
• Objectives
Observation
– Environment and uncertain reasoning
– Intuitive comparison of alternative knowledge
representation (KR) and inference formalisms Action
Logic, fuzzy set, certainty factor, & probability
Uncertain knowledge and dependency • Ex Roles of doctor, tutor, translator, technician,
Belief retraction and bi-directional inference robot, detective, ..., in their environment
– Formal justification of Bayesian approach • Reasoning tasks of agent
• References – What is the state of the environment?
– Pearl (1.1, 1.2) – What is the next action to take?
– Russell & Norvig (Chap.13, Section 14.7) • Terms: reasoning, inference, decision making
Y. Xiang, CIS 6120, Uncertian Reasoning 1 Y. Xiang, CIS 6120, Uncertian Reasoning 2

Partially Observable Environment Stochastic Environment

• Examples
– Image vs object • Ex Cavity vs. toothache
– Sensor output vs sensed • Ex Dead battery vs. non-starting car
– Speech vs spoken words • When relations btw events are stochastic, how can
– Symptom vs disease
agent determine the unobservable?
– Crime scene vs criminal
– Sentence vs meaning • Ex Troubleshooting complex equipment
– Current vs future events • Even if env is deterministic in theory, it is often too
• Environment is often partially observable. costly to observe everything relevant.
– Not everything of interest is observable. – The env practically becomes partially observable
– Observations are imprecise, partial and noisy. and stochastic.

Y. Xiang, CIS 6120, Uncertian Reasoning 3 Y. Xiang, CIS 6120, Uncertian Reasoning 4

1
Uncertain Reasoning

Comparison Of Alternative Formalisms Ex Rule-Based Expert System On Plant Family

• We compare, intuitively, alternative KR & inference • The knowledge base

formalisms and how they behave in uncertain env. 1. IF class is gymnosperm AND leaf shape is scale-like,
THEN family is cypress.
• Mathematical logic 2. IF class is gymnosperm AND leaf shape is needle-like
– What is known is stated as a set PR of AND pattern is random, THEN family is pine.
consistent premises. 3. IF type is tree AND outlook is not broad and flat,
THEN class is gymnosperm.
– A statement not in PR can only be accepted if it 4. IF stem is woody AND position is creeping
is derivable from PR by sound rules of inference. AND there is one main trunk, THEN type is tree.
– Ex Rule-based expert systems
– Logic approach dominated AI in 1960s and 70s.

Y. Xiang, CIS 6120, Uncertian Reasoning 5 Y. Xiang, CIS 6120, Uncertian Reasoning 6

Ex Rule-Based System (cont) Represent Uncertain Knowledge In Logic

• Observations over plant X • Ex Diagnosis in internal medicine.

– X’s leaf shape is scale-like. – Is the following representation correct?
– X’s outlook is not broad and flat. p Cough(p)  Pneumonia(p)
– X’s stem is woody AND position is creeping AND there is
one main trunk. – Can the following representation fix the problem?
• The query p Pneumonia(p)  Cough(p).
– Is the family of plant X cypress? – Still another alternative
• How would this piece of knowledge be expressed
• Inference by backward chaining
according to probability theory?

Y. Xiang, CIS 6120, Uncertian Reasoning 7 Y. Xiang, CIS 6120, Uncertian Reasoning 8

2
Uncertain Reasoning

Decision Making In Logic Faithful Representation Of Dependency

• When a logic agent knows enough about the env,
it can derive plans that guarantee to work. • To infer unobservable from observable, info from
multiple sources should be combined without bias.
– What if it does not have all needed facts?
• Ex There’re at most 2 obstacles in a grid world. • Ex A murder suspect was seen on site but no
Robi must deliver from (1,1) to (9,9). At cells (1,2) motivation can be established.
and (2,1), it detects obstacles nearby.
– What should be Robi’s plan? • Information from the same source should not be
• Ex Agent’ goal: print a file to a networked printer. double-counted.
– What should be its plan? • Ex Coordinated reviewers
• A logic agent may not be able to act.

Y. Xiang, CIS 6120, Uncertian Reasoning 9 Y. Xiang, CIS 6120, Uncertian Reasoning 10

Rule-Based System With Certainty Factors

Fuzzy Set And Fuzzy Logic (Zadeh, 1965)
(Shortliffe, 76)
• A fuzzy sentence has a truth value in [0,1].
• To manage uncertainty in MICIN, the rule-based
– Ex T(tall(David)) = 0.7 system was enhanced with certainty factors.
• Fuzzy truth of a complex sentence
• Ex Fever, headache and flu
– T(a  b) = min ( T(a), T(b) )
– If fever, then flu with 0.7 certainty.
– T(a  b) = max ( T(a), T(b) )
– If headache, then flu with 0.6 certainty.
– T(a) = 1 - T(a)
• Ex T( head(Coin_toss)  tail(Coin_toss) ) = ? – If fever and headache, what’s the CF for flu?
– What’s wrong? • What would the system conclude with coordinated
– What is the probability of the joint event? reviewers?
• Alternative views on fuzzy logic
Y. Xiang, CIS 6120, Uncertian Reasoning 11 Y. Xiang, CIS 6120, Uncertian Reasoning 12

3
Uncertain Reasoning

Retraction Of Uncertain Conclusion Bi-directional Inference

• A piece of knowledge should support both forward
• Retraction is common in uncertain reasoning. and backward inference.
– Ex Press power button, but stereo is not on. No – Ex Contaminated oxygen may cause fighter pilot
to hallucinate.
electricity?
• Combination of forward and backward inference
– Ex Lottery ticket enables explaining away.
• Does logic allow retraction? – Ex Contaminated oxygen as well as weakened
• Default logic (Reiter, 80), non-monotonic logics health may cause fighter pilot to hallucinate.
(McDermott, 80), and circumscription (McCarthy, 80) Observing hallucination, what do you believe
• Does probability allow retraction? to be the cause?
Upon seeing a recent health report, what is
your updated belief?
Y. Xiang, CIS 6120, Uncertian Reasoning 13 Y. Xiang, CIS 6120, Uncertian Reasoning 14

Formal Justification Of Bayesian Probability

Bi-directional Inference: Rule Vs. Probability
(Cox, 1946)
• Rule-based system • Interpretations of probability
– A rule can be used only in one way. – Frequentist, classical, and Bayesian
Rules are justified by Modus Ponens. • Cox’s assumptions (appealing to common sense)
– Coexistence of forward and backward rules 1. Propositional logic.
causes circular inference. Consequence? 2. Likelihood c.b|a = F(c|b.a,b|a).
• Probability 3. Likelihood b|b.a = 1.
– P(e|c) supports both forward and backward 4. Likelihood ~b|a = S(b|a).
inference. • From these, Cox derived axioms of probability
– Bayes’ rule: P(c|e) = P(e|c)P(c)/P(e) theory.
– We will show how probability allows explaining
away.
Y. Xiang, CIS 6120, Uncertian Reasoning 15 Y. Xiang, CIS 6120, Uncertian Reasoning 16

4
Uncertain Reasoning

Review Of Propositional Logic Review Of Propositional Logic

• Logical Equivalences
• BNF grammar for sentences
Commutativity: (  ) = (  ), (  ) = (  )
Symbol  a | b | cancer | alarm
Associativity: ((  )  ) = (  (   ))
AtomicSentence  true | false | Symbol
((  )  ) = (  (   ))
Sentence  AtomicSentence | ComplexSentence
Distributivity: ((  )  ) = ((  )  (   ))
ComplexSentence   Sentence
((  )  ) = ((  )  (   ))
| (Sentence  Sentence)
Double-negation elimination: () = 
| (Sentence  Sentence)
Contraposition: (  ) = (  )
| (Sentence  Sentence) Implication elimination: (  ) = (  )
| (Sentence  Sentence)
Biconditional elimination:(  ) = ((  )  (  ))
De Morgan: (  ) = (  ), (  ) = (  )
Y. Xiang, CIS 6120, Uncertian Reasoning 17 Y. Xiang, CIS 6120, Uncertian Reasoning 18

Justification By Connecting Belief And Action

Axioms Of Probability
(Bruno de Finetti ,1937)
• Notation: Let a and b denote distinct propositions. • Ag’s degree of belief in e being p implies that
• Range: 0  P(a)  1. – Ag is indifferent btw lotteries L1 = {$1|e, $0|e}
• Certainty: P(true) = 1. and L2 = {$p|}. Why?
• Sum: Whenever a and b are mutually exclusive, – Ag is indifferent among L3 = {$(p-1)|e, $p|e},
P(ab) = P(a) + P(b). L4 = {$(1-p)|e, -$p|e}, and L5 = {$0|}. Why?
• Product: P(ab) = P(a|b) P(b). • Either Ag’s degree of belief follows axioms of
• Note probability (hence it is probability), or it does not.
– From axioms, the rest of probability theory is • There exists a combination lottery that guarantees
derivable.
Ag’s loss iff its degree of belief is not probability.
– The set of axioms is not unique.
Y. Xiang, CIS 6120, Uncertian Reasoning 19 Y. Xiang, CIS 6120, Uncertian Reasoning 20

5
Uncertain Reasoning

What Happens When Belief Is Not Probability What Happens With Probabilistic Belief

ab ab ab ab

Proposition Belief Lottery
ab ab ab ab
Proposition Belief Lottery
a 0.1 {-$0.9|a, -$0.9 -$0.9 $0.1 $0.1
a 0.1 {-$0.9|a, -$0.9 -$0.9 $0.1 $0.1
$0.1|a}
$0.1|a}
b 0.6 {-$0.4|b, -$0.4 $0.6 -$0.4 $0.6
$0.6|b} b 0.6 {-$0.4|b, -$0.4 $0.6 -$0.4 $0.6
ab 0.8 {$0.2|ab, $0.2 $0.2 $0.2 -$0.8 $0.6|b}
-$0.8|(ab)}
ab 0.7 {$0.3|ab, $0.3 $0.3 $0.3 -$0.7
Combo: -$1.1 -$0.1 -$0.1 -$0.1 -$0.7|(ab)}

Combo: - $1.0 $0 $0 $0

Y. Xiang, CIS 6120, Uncertian Reasoning 21 Y. Xiang, CIS 6120, Uncertian Reasoning 22

What Happens With Probabilistic Belief What Have Been Shown?

• How about a different lottery? • An agent acting in an uncertain environment faces
decision situations like lotteries all the time.
ab ab ab ab
Proposition Belief Lottery
• If an agent does not base its belief on probability, it
a 0.1 {-$0.9|a, -$0.9 -$0.9 $0.1 $0.1
$0.1|a}
will encounter decision situations (by nature or by
design) where its performance is guaranteed to be
b 0.6 {-$0.4|b, -$0.4 $0.6 -$0.4 $0.6
$0.6|b} sub-optimal.
ab 0.7 {-$0.3|ab, -$0.3 -$0.3 -$0.3 $0.7 • The only way that the agent can always avoid such
$0.7|(ab)} undesirable position is to adopt probabilistic belief.
Combo: - $ 1.6 - $ 0.6 - $ 0.6 $1.4

Y. Xiang, CIS 6120, Uncertian Reasoning 23 Y. Xiang, CIS 6120, Uncertian Reasoning 24