Reasoning in Uncertain Situation

1
Reasoning in Uncertain Situation

• Chap. 5, Chap. 9

2
Managing Uncertainty

• Abduction
• If P ? Q, Q then P ? Not sound
• Most diagnostic rules are abduction
• True implication
• Problem Symptom
• Exgt Problem in battery Light gone
• Expert system rule
• Symptom Problem
• Exgt Light gone Problem in battery
• ? Need measure of uncertainty

3
Certainty Factor

• Certainty Theory
• Confidence measure (expert heuristic) combining
  rules
• Certainty Factor(CF) of H given E (E ? H)
• CF(H E) MB(H E) - MD(H E)
• MB Measure of belief of hypothesis H given
  evidence E
• MD Measure of disbelief of hypothesis H given
  evidence E
• Inference with CF
• Rule if P then Q (CF 0.9)
• Fact P (CF 0.9) Q (CF 0.90.9
  0.81)
• Combining CF
• CF(P1 and P2) min (CF(P1), CF(P2))
• CF(P1 or P2) max (CF(P1), CF(P2))

4
Certainty Factor

• Example
• Rule (P1 and P2) or P3 R1(0.7) and R2(0.3)
• During production cycle, it is found that
• CF(P1) 0.6
• CF(P2) 0.4
• CF(P3) 0.2
• Then
• CF((P1 and P2) or P3) max (min (0.6, 0.4),
  0.2)
• 0.4
• \CF(R1) 0.4 0.7 0.28
• CF(R2) 0.4 0.3 0.12

5
MYCIN

• Purpose
• Expert system determines the organism that
  causes meningitis and provides therapy
• Inference
• Start from a hypothesis(organism), perform
  goal-driven search
• Use certainty factor
• Example rule
• If ((infection is primary-bacteremia) AND
• (portal-of-entry is gastrointestinal-tract)
  )
• Then (conclude bacteroid 0.7)

6
Reasoning with Fuzzy Sets

• Fuzzy set
• Set membership is a function to 0, 1
• Exgt S set of small integer
• fs(1) 1.0, fs(5) 0.7, fs(100) 0.001
• Membership function
• Sets S, M, T

f(x)
S
M
T
1
x


160
170
180
fT(178) 0.7, fM(178) 0.3
7
Reasoning with Fuzzy Sets

• Fuzzy logic
• If fA(x) a, fB(x) b then
• fAÇB(x) min (a, b)
• fAÈB(x) max (a, b)
• Fuzzy rules
• high(speed) and many(cars) push(brake)
• very_high(temp) cut(power)

8
Reasoning with Fuzzy Sets

• Example Inverted Pendulum
• Input x1 ?, x2 d?/dt
• Output u movement M

9
Reasoning with Fuzzy Sets

• Fuzzy set
• x1, x2 Z(zero), P(positive), N(negative)
• u Z, P, PB(positive big), N, NB(negative big)
• Rules
• If x1 P and x2 Z then u P
• If x1 P and x2 N then u Z
• If x1 Z and x2 Z then u Z
• If x1 Z and x2 N then u N
• . . .

10
Reasoning with Fuzzy Sets
11
Reasoning with Fuzzy Sets

• If x1 ? 1, x2 d?/dt -4
• 1. from fuzzy set membership,
• x1 0.5P, 0.5Z
• x2 0.8N, 0.2Z

12
Reasoning with Fuzzy Sets

• 2. From fuzzy rules,
• u 0.2P
• (If x1 P(0.5) and x2 Z(0.2)
• then u P(min(0.5,0.2)),
• u 0.5Z
• (If x1 P(0.5) and x2 N(0.8)
• then u Z(min(0.5,0.2)),
• u 0.2Z
• (If x1 P(0.5) and x2 Z(0.2)
• then u Z(min(0.5,0.2)),
• u 0.5N
• (If x1 Z(0.5) and x2 N(0.8)
• then u N(min(0.5,0.8)),

13
Reasoning with Fuzzy Sets

• 3. From fuzzy set membership,
• u -2

Fuzzification
X1 1, X2 -4
P, Z, N
Fuzzy rules
Defuzzification
U -2
P, Z, N
14
Probability

• For an experiment that produces outcomes
• Sample space S set of all outcomes
• Event E a subset of S
• Probability P(E) E/S (if outcomes are
  equally likely)
• Probability distribution
• Function P x ? 0, 1 (x random variable)

15
Probability

• Joint probability distribution
• P(x,y) represents joint probability
  distribution of x and y
• Example
• Gastrue,false, Meterempty, full, Startyes,
  no
• P(G,M,S)

16
Probability

• P(Gasfalse, Meterempty, Startno) 0.1386
• P(Startyes) 0.5620
• Prior probability
• P(Startyes), given that Meterempty ?
• Posterior probability (conditional probability)

17
Conditional Probability

• Conditional probability
• P(AB) P(A Ù B) / P(B)
• Example
• P(A Ù B) 4 / 100 0.04
• P(B) 5 / 100 0.05
• P(A B) 4 / 5 0.80
• Independence
• P(AB) P(A)
• P(A Ù B) P(AB) P(B)
• P(A) P(B) if A, B are independent

18
Bayesian Reasoning

• Let H Hypothesis (Problem)
• E Evidence (Symptom)
• Then the probability of E ? H is P(H E)
• P(H Ç E) P(H E) P(E)
• P(E H) P(H)
• ? P(H E) P(E H) P(H)
• P(E)

19
Bayesian Reasoning

• If we want to compare P(H1 E), P(H2 E),
• P(H1 E) P(E H1) P(H1)
• P(E)
• P(H2 E) P(E H2) P(H2)
• P(E)
• ? P(Hi E) ? P(E Hi) P(Hi)

same
20
Bayesian Reasoning

• For multiple, independent evidences
• P(E1, E2, En) P(E1)P(E2) P(En)
• ? P(Hi E1, E2, En)
• ? P(E1, E2, En Hi) P(Hi)
• ? P(E1 Hi) P(E2 Hi) P(E1 Hi)
  P(Hi)

21
Bayesian Classifier Example? Example ?? (hw3)

• Category C1 of sample documents 60
• Keywords - information (2), network (12),
  algorithm (6), system (10)
• Category C2 of sample documents 40
• Keywords - information (10), database (8),
  algorithm (2), system (10)
• New document D has keywords - information,
  system
• P(C1 information, system)
• k P(information C1) P(system C1)
  P(C1)
• k 2/30 10/30 60/100 0.013 k
• P(C2 information, system)
• k P(information C2) P(system C2)
  P(C2)
• k 10/30 10/30 40/100 0.044 k
• D is classified to C2

22
Bayesian Reasoning with Full Probability
Distribution

• P(Gasfalse, Meterempty, Startno) 0.1386
• P(Gasfalse) 0.2
• P(Startyes Meterfull) P(Syes, Mfull) /
  P(Mfull)
• (0.50400.0006) / (0.50400.00060.21600.0594
  )
• 0.6469

23
Bayesian Reasoning with Full Probability
Distribution

• P(Startyes Meterfull)

Select Mfull
Sum G
Normalize
24
Bayesian Belief Network

• Nodes Random variables
• Edges Direct influence
• Each node x stores P(x parents(x))
• P(G,M,S) P(G) P(M G) P(S G)

25
Independency (d-separation)

• A, B are independent
• P(BA) P(B)
• A, B are conditionally dependent
• P(BA,C) ? P(BC)
• A, B are dependent
• P(BA) ? P(B)
• A, B are conditionally independent
• P(BA,C) P(BC)

26
Independency (d-separation)

• Gas and Start are dependent
• Gas and Plug are independent
• Gas and Plug are conditionally dependent given
  Start
• Meter and Start are conditionally independent
  given Gas
• P(S M, G) P(S G)

27
Chain Rule and Independence

• P(A,B) P(A,B) P(A) P(BA) P(A)
• P(A)
• P(A,B,C) P(A,B,C) P(A,B) P(A) P(CB,A) P(BA)
  P(A)
• P(A,B) P(A)
• P(A,B,C,D) P(DC,B,A) P(CB,A) P(BA) P(A)
• P(G,M,S) P(S G,M) P(M G) P(G) (Chain
  rule)
• P(S G) P(M G) P(G)
• (If S, M are conditionally independent given
  G)
• P(S) P(M) P(G)
• (If S, M, G are all independent)

28
Joint Distribution from Bayesian Network

• In general,
• Joint probability ? product of conditional
  probabilities
• P(A,B,C,D,E) P(A)P(BA)P(CA,B)P(DA,B,C)P(EA,B
  ,C,D)
• P(A)P(BA)P(CB) P(DB,C) P(EC,D)

p(EC,D)
p(DB,C)
p(CB)
p(A)
p(BA)
29
Reasoning with Bayesian Belief Network

• Inference for P(H E)
• From the product of probability table,
• Remove all rows except E
• Compute product
• Sum over irrelevant variables
• Normalize
• Example
• P(Syes Mfull)

30
Reasoning with Bayesian Belief Network

p(S,G,M)
Remove Mempty
Product
p(S,G Mf)
Sum over G
p(S Mf)
Normalize
p(S Mf)
31
Reasoning with Bayesian Belief Network
32
Reasoning with Bayesian Belief Network

• Advantage
• Assume 1 H, 30 E
• Compute P(H E1, E2, , E30)
• P(E1, E2, E30 H) P(H)
• P(E1, E2, , E30)
• From
• P(H, E1, E2, E30)
• ? 231 2,147,483,648 prob.
• Or from
• P(E1) P(E2) P(E3 E1, E2)
• (assume less than 2 parents in Bayesian Network)
• ? less than 831 248 prob.

33
Probabilistic FSM

• Finite state machine where
• Next state function is a probability distribution

34
Markov Model

• Markov process (Markov chain)
• Probability of a state at time t depends on its
  previous n states
• P(?t ?t-1 , ?t-2 , ?t-3 , , ?t-n )
• First-order Markov process
• Probability of a state at time t depends on its
  previous 1 state
• P(?t ?t-1 )

35
Markov Model

• Example
• 4 states S1(sunny), S2(cloudy), S3(Foggy),
  S4(Rainy)
• Today is sunny. Prob. of next 2 days are rainy?
• P(S1, S4, S4) P(S1) P(S4 S1) P(S4 S1,S4)
• P(S1) P(S4 S1) P(S4 S4)
• 1 0.1 0.2 0.02

36
Hidden Markov Model (HMM)

• HMM
• States are hidden
• Probability of observation is given. P(Oj Si)

37
Hidden Markov Model (HMM)

• P(s1 sn o1 on)
• P(o1 on s1 sn) P(s1 sn)
• P(o1 on)
• a P(o1 on s1 sn) P(s1 sn)
• a P(o1 s1)P(o2 s2)P(on sn) P(s1 sn)
• a P(o1 s1)P(o2 s2)P(on sn) P(s1)P(s2
  s1)P(sn sn-1)
• ? i1..n P(oi si ) P(si si-1)

38
Speech Recognition

• The problem
• Observed sequence of acoustic signals
• lt, n, iy, gt
• Determine Which word ?
• ltneedgt, ltkneegt, ltnewgt?
• Compute P(word signal) by using HMM (Viterbi
  algorithm)
• P(w1, w2, w3 o1, o2, o3) ?
• P(w4, w5, w6 o1, o2, o3) ?
• P(w7, w8, w9 o1, o2, o3) ?
• Find max P(s1 sn o1 on)

39
Speech Recognition

• max P(s1 sn o1 on)
• max P(o1 on s1 sn) P(s1 sn)
• P(o1 on)
• max a P(o1 s1) P(o2 s2) P(on sn)
• P(s1) P(s2 s1) P(sn sn-1)
• max ? i1..n P(oi si ) P(si si-1)

40
Speech Recognition
state (hidden)
observation
41
Handwriting Recognition

• Hand-written character recognition

a or 6 ?
5 or S ?
42
Handwriting Recognition

• The problem
• Observed output sequence of moving directions
  d1.. dn
• Find the sequence of states s1.. sn for each
  character
• that maximize the probability
• P(s1 sn d1 dn ) ? i1..n P(di si )
  P(si si-1)

43
Handwriting Recognition

• Example
• Writing 8, 8, 7, 7, 7, 6, 6, 5, 5,
• P(States of zero 8, 8, 7, 7, 7, 6, 6, 5, 5,
  ) gtgt
• P(States of one 8, 8, 7, 7, 7, 6, 6, 5, 5,
  )
• Markov process
• Output process