Sampling Techniques for Probabilistic and Deterministic Graphical - - PowerPoint PPT Presentation

Sampling Techniques for Probabilistic and Deterministic Graphical models

ICS 276, Spring 2017 Bozhena Bidyuk Rina Dechter

Reading” Darwiche chapter 15, related papers

Overview

• 1. Probabilistic Reasoning/Graphical models
• 2. Importance Sampling
• 3. Markov Chain Monte Carlo: Gibbs Sampling
• 4. Sampling in presence of Determinism
• 5. Rao-Blackwellisation
• 6. AND/OR importance sampling

Markov Chain

• A Markov chain is a discrete random process with

the property that the next state depends only on the current state (Markov Property):

3

x1 x2 x3 x4

) | ( ) ,..., , | (

1 1 2 1  



t t t t

x x P x x x x P

• If P(Xt|xt-1) does not depend on t (time

homogeneous) and state space is finite, then it is

• ften expressed as a transition function (aka

transition matrix)

1 ) (  



x

x X P

Example: Drunkard’s Walk

• a random walk on the number line where, at

each step, the position may change by +1 or −1 with equal probability

4

1 2 3

,...} 2 , 1 , { ) (  X D

5 . 5 . ) 1 ( ) 1 ( n n P n P  

transition matrix P(X) 1 2

Example: Weather Model

5

5 . 5 . 1 . 9 . ) ( ) ( sunny rainy sunny P rainy P

transition matrix P(X)

} , { ) ( sunny rainy X D 

rain rain rain rain sun

Multi-Variable System

• state is an assignment of values to all the

variables

6

x1

t

x2

t

x3

t

x1

t+1

x2

t+1

x3

t+1

} ,..., , {

2 1 t n t t t

x x x x  finite discrete X D X X X X

i

, ) ( }, , , {

3 2 1

 

Bayesian Network System

• Bayesian Network is a representation of the

joint probability distribution over 2 or more variables

X1

t

X2

t

X3

t

} , , {

3 2 1 t t t t

x x x x 

X1 X2 X3

7

} , , {

3 2 1

X X X X 

x1

t+1

x2

t+1

x3

t+1

8

Stationary Distribution Existence

• If the Markov chain is time-homogeneous,

then the vector (X) is a stationary distribution (aka invariant or equilibrium distribution, aka “fixed point”), if its entries sum up to 1 and satisfy:

• Finite state space Markov chain has a unique

stationary distribution if and only if:

– The chain is irreducible – All of its states are positive recurrent







) (

) | ( ) ( ) (

X D x j i j i

i

x x P x x  

9

Irreducible

• A state x is irreducible if under the transition rule
• ne has nonzero probability of moving from x to

any other state and then coming back in a finite number of steps

• If one state is irreducible, then all the states

must be irreducible

(Liu, Ch. 12, pp. 249, Def. 12.1.1)

10

Recurrent

• A state x is recurrent if the chain returns to x

with probability 1

• Let M(x ) be the expected number of steps to

return to state x

• State x is positive recurrent if M(x ) is finite

The recurrent states in a finite state chain are positive recurrent .

Stationary Distribution Convergence

• Consider infinite Markov chain:

n n n

P P x x P P

) (

) | (  

• Initial state is not important in the limit

“The most useful feature of a “good” Markov chain is its fast forgetfulness of its past…” (Liu, Ch. 12.1)

) (

lim

n n

P

 

 

11

• If the chain is both irreducible and aperiodic,

then:

Aperiodic

• Define d(i) = g.c.d.{n > 0 | it is possible to go

from i to i in n steps}. Here, g.c.d. means the greatest common divisor of the integers in the

• set. If d(i)=1 for i, then chain is aperiodic
• Positive recurrent, aperiodic states are ergodic

12

Markov Chain Monte Carlo

• How do we estimate P(X), e.g., P(X|e) ?

13

• Generate samples that form Markov Chain

with stationary distribution =P(X|e)

• Estimate  from samples (observed states):

visited states x0,…,xn can be viewed as “samples” from distribution 

) , ( 1 ) (

1 t T t

x x T x





   ) ( lim x

T

 

 



MCMC Summary

• Convergence is guaranteed in the limit
• Samples are dependent, not i.i.d.
• Convergence (mixing rate) may be slow
• The stronger correlation between states, the

slower convergence!

• Initial state is not important, but… typically,

we throw away first K samples - “burn-in”

14

Gibbs Sampling (Geman&Geman,1984)

• Gibbs sampler is an algorithm to generate a

sequence of samples from the joint probability distribution of two or more random variables

• Sample new variable value one variable at a

time from the variable’s conditional distribution:

• Samples form a Markov chain with stationary

distribution P(X|e)

15

) \ | ( } ,..., , ,.., | ( ) (

1 1 1 i t i t n t i t i t i i

x x X P x x x x X P X P  

 

Gibbs Sampling: Illustration

The process of Gibbs sampling can be understood as a random walk in the space of all instantiations of X=x (remember drunkard’s walk): In one step we can reach instantiations that differ from current one by value assignment to at most one variable (assume randomized choice of variables Xi).

Ordered Gibbs Sampler

Generate sample xt+1 from xt : In short, for i=1 to N:

17

) , \ | ( from sampled ) , ,..., , | ( ... ) , ,..., , | ( ) , ,..., , | (

1 1 1 1 2 1 1 1 3 1 1 2 1 2 2 3 2 1 1 1 1

e x x X P x X e x x x X P x X e x x x X P x X e x x x X P x X

i t i t i i t N t t N t N N t N t t t t N t t t

       

        

Process All Variables In Some Order

Transition Probabilities in BN

Markov blanket:

: ) | ( ) \ | (

t i i i t i

markov X P x x X P 







i j ch

X j j i i i t i

pa x P pa x P x x x P ) | ( ) | ( ) \ | ( ) ( ) (



 

j j ch

X j i i i

pa ch pa X markov





Xi

Given Markov blanket (parents, children, and their parents), Xi is independent of all other nodes

18

Computation is linear in the size of Markov blanket!

Ordered Gibbs Sampling Algorithm (Pearl,1988)

Input: X, E=e Output: T samples {xt} Fix evidence E=e, initialize x0 at random

1. For t = 1 to T (compute samples) 2. For i = 1 to N (loop through variables) 3. xi

t+1  P(Xi | markovi t)

4. End For 5. End For

Gibbs Sampling Example - BN

20 X1 X4 X8 X5 X2 X3 X9 X7 X6

} { }, ,..., , {

9 9 2 1

X E X X X X  

X1 = x1 X6 = x6 X2 = x2 X7 = x7 X3 = x3 X8 = x8 X4 = x4 X5 = x5

Gibbs Sampling Example - BN

21 X1 X4 X8 X5 X2 X3 X9 X7 X6

) , ,..., | (

9 8 2 1 1 1

x x x X P x 

} { }, ,..., , {

9 9 2 1

X E X X X X  

) , ,..., | (

9 8 1 1 2 1 2

x x x X P x 



Answering Queries P(xi |e) = ?

• Method 1: count # of samples where Xi = xi (histogram estimator):





  

T t t i i i i i

markov x X P T x X P

1

) | ( 1 ) (





 

T t t i i i

x x T x X P

1

) , ( 1 ) ( 

Dirac delta f-n

• Method 2: average probability (mixture estimator):
• Mixture estimator converges faster (consider

estimates for the unobserved values of Xi; prove via Rao-Blackwell theorem)

Rao-Blackwell Theorem

Rao-Blackwell Theorem: Let random variable set X be composed of two groups of variables, R and L. Then, for the joint distribution (R,L) and function g, the following result applies

23

)] ( [ } | ) ( { [ R g Var L R g E Var 

for a function of interest g, e.g., the mean or covariance (Casella&Robert,1996, Liu et. al. 1995).

• theorem makes a weak promise, but works well in practice!
• improvement depends the choice of R and L

Importance vs. Gibbs

 

   

       

T t t t t t T t t T t

x Q x P x g T g e X Q X x g T X g e X P e X P e X P x

1 1

) ( ) ( ) ( 1 ) | ( ) ( 1 ) ( ˆ ) | ( ) | ( ˆ ) | ( ˆ

wt

Gibbs: Importance:

Gibbs Sampling: Convergence

25

• Sample from `P(X|e)P(X|e)
• Converges iff chain is irreducible and ergodic
• Intuition - must be able to explore all states:

– if Xi and Xj are strongly correlated, Xi=0 Xj=0, then, we cannot explore states with Xi=1 and Xj=1

• All conditions are satisfied when all

probabilities are positive

• Convergence rate can be characterized by the

second eigen-value of transition matrix

Gibbs: Speeding Convergence

Reduce dependence between samples (autocorrelation)

• Skip samples
• Randomize Variable Sampling Order
• Employ blocking (grouping)
• Multiple chains

Reduce variance (cover in the next section)

26

Blocking Gibbs Sampler

• Sample several variables together, as a block
• Example: Given three variables X,Y,Z, with domains of

size 2, group Y and Z together to form a variable W={Y,Z} with domain size 4. Then, given sample (xt,yt,zt), compute next sample: + Can improve convergence greatly when two variables are strongly correlated!

• Domain of the block variable grows exponentially with

the #variables in a block!

27

) | , ( ) , ( ) ( ) , | (

1 1 1 1 1     

   

t t t t t t t t

x Z Y P w z y w P z y X P x

Gibbs: Multiple Chains

• Generate M chains of size K
• Each chain produces independent estimate Pm:

28

   

• 
• 

 M i m

P M P

1

1 ˆ







K t i t i i m

x x x P K e x P

1

) \ | ( 1 ) | (

Treat Pm as independent random variables.

• Estimate P(xi|e) as average of Pm (xi|e) :

Gibbs Sampling Summary

• Markov Chain Monte Carlo method

(Gelfand and Smith, 1990, Smith and Roberts, 1993, Tierney, 1994)

• Samples are dependent, form Markov Chain
• Sample from

which converges to

• Guaranteed to converge when all P > 0
• Methods to improve convergence:

– Blocking – Rao-Blackwellised

29

) | ( e X P ) | ( e X P

Overview

• 1. Probabilistic Reasoning/Graphical models
• 2. Importance Sampling
• 3. Markov Chain Monte Carlo: Gibbs Sampling
• 4. Sampling in presence of Determinism
• 5. Rao-Blackwellisation
• 6. AND/OR importance sampling

Sampling: Performance

• Gibbs sampling

– Reduce dependence between samples

• Importance sampling

– Reduce variance

• Achieve both by sampling a subset of variables

and integrating out the rest (reduce dimensionality), aka Rao-Blackwellisation

• Exploit graph structure to manage the extra cost

31

Smaller Subset State-Space

• Smaller state-space is easier to cover

32

} , , , {

4 3 2 1

X X X X X  } , {

2 1 X

X X  64 ) (  X D 16 ) (  X D

Smoother Distribution

33

00 01 10 11 0.1 0.2 00 01 10 11

P(X1,X2,X3,X4)

0-0.1 0.1-0.2 0.2-0.26 1 0.1 0.2 1

P(X1,X2)

0-0.1 0.1-0.2 0.2-0.26

Speeding Up Convergence

• Mean Squared Error of the estimator:

   

P Var BIAS P MSE

Q Q

 

2

 

        

2 2

] [ ˆ ] ˆ [ ] ˆ [ P E P E P Var P MSE

Q Q Q Q

• Reduce variance  speed up convergence !
• In case of unbiased estimator, BIAS=0

34

Rao-Blackwellisation

35

)} ( ~ { ]} | ) ( [ { )} ( { )} ( ˆ { ]} | ) ( [ { )} ( { ]} | ) ( {var[ ]} | ) ( [ { )} ( { ]} | ) ( [ ] | ) ( [ { 1 ) ( ~ )} ( ) ( { 1 ) ( ˆ

1 1

x g Var T l x h E Var T x h Var x g Var l x g E Var x g Var l x g E l x g E Var x g Var l x h E l x h E T x g x h x h T x g L R X

T T

              



Liu, Ch.2.3

Rao-Blackwellisation

• X=RL
• Importance Sampling:
• Gibbs Sampling:

– autocovariances are lower (less correlation between samples) – if Xi and Xj are strongly correlated, Xi=0  Xj=0,

• nly include one fo them into a sampling set

36

} ) ( ) ( { } ) , ( ) , ( { R Q R P Var L R Q L R P Var

Q Q



Liu, Ch.2.5.5 “Carry out analytical computation as much as possible” - Liu

Blocking Gibbs Sampler vs. Collapsed

• Standard Gibbs:

(1)

• Blocking:

(2)

• Collapsed:

(3)

37

X Y Z

) , | ( ), , | ( ), , | ( y x z P z x y P z y x P ) | , ( ), , | ( x z y P z y x P ) | ( ), | ( x y P y x P

Faster Convergence

Collapsed Gibbs Sampling

Generating Samples

Generate sample ct+1 from ct :

38

) , \ | ( ) , ,..., , | ( ... ) , ,..., , | ( ) , ,..., , | (

1 1 1 1 2 1 1 1 3 1 1 2 1 2 2 3 2 1 1 1 1

e c c c P c C e c c c c P c C e c c c c P c C e c c c c P c C

i t i t i i t K t t K t K K t K t t t t K t t t

from sampled        

        

In short, for i=1 to K:

Collapsed Gibbs Sampler

Input: C X, E=e Output: T samples {ct } Fix evidence E=e, initialize c0 at random

1. For t = 1 to T (compute samples) 2. For i = 1 to N (loop through variables) 3. ci

t+1  P(Ci | ct\ci)

4. End For 5. End For

Calculation Time

• Computing P(ci| ct\ci,e) is more expensive

(requires inference)

• Trading #samples for smaller variance:

– generate more samples with higher covariance – generate fewer samples with lower covariance

• Must control the time spent computing

sampling probabilities in order to be time- effective!

40

Exploiting Graph Properties

Recall… computation time is exponential in the adjusted induced width of a graph

• w-cutset is a subset of variable s.t. when they

are observed, induced width of the graph is w

• when sampled variables form a w-cutset ,

inference is exp(w) (e.g., using Bucket Tree Elimination)

• cycle-cutset is a special case of w-cutset

41

Sampling w-cutset  w-cutset sampling!

What If C=Cycle-Cutset ?

42

} { }, {

9 5 2

X E ,x x c  

X1 X7 X5 X4 X2 X9 X8 X3 X6 X1 X7 X4 X9 X8 X3 X6

P(x2,x5,x9) – can compute using Bucket Elimination P(x2,x5,x9) – computation complexity is O(N)

Computing Transition Probabilities

43

) , , 1 ( : ) , , ( :

9 3 2 9 3 2

x x x P BE x x x P BE  

X1 X7 X5 X4 X2 X9 X8 X3 X6

) , , 1 ( ) | 1 ( ) , , ( ) | ( ) , , 1 ( ) , , (

9 3 2 3 2 9 3 2 3 2 9 3 2 9 3 2

x x x P x x P x x x P x x P x x x P x x x P             

Compute joint probabilities: Normalize:

Cutset Sampling-Answering Queries

• Query: ci C, P(ci |e)=? same as Gibbs:

44

computed while generating sample t using bucket tree elimination compute after generating sample t using bucket tree elimination







T t i t i i

e c c c P T |e c P

1

) , \ | ( 1 ) ( ˆ







T t t i i

,e c x P T |e) (x P

1

) | ( 1

• Query: xi X\C, P(xi |e)=?

Cutset Sampling vs. Cutset Conditioning

45

) | ( ) | ( ) ( ) | ( ) | ( 1

) ( ) ( 1

e c P c,e x P T c count c,e x P ,e c x P T |e) (x P

C D c i C D c i T t t i i

    

  

  

• Cutset Conditioning
• Cutset Sampling

) | ( ) | (

) (

e c P c,e x P |e) P(x

C D c i i

  



Cutset Sampling Example

46

                ) ( ) ( ) ( 3 1 ) | ( ) ( ) ( ) (

9 2 5 2 9 1 5 2 9 5 2 9 2 9 2 5 2 3 2 9 1 5 2 2 2 9 5 2 1 2

,x | x x P ,x | x x P ,x | x x P x x P ,x | x x P x ,x | x x P x ,x | x x P x  

X1 X7 X6 X5 X4 X2 X9 X8 X3

Estimating P(x2|e) for sampling node X2 :

Sample 1 Sample 2 Sample 3

Cutset Sampling Example

47

) , , | ( } , { ) , , | ( } , { ) , , | ( } , {

9 3 5 3 2 3 3 5 3 2 3 9 2 5 2 2 3 2 5 2 2 2 9 1 5 1 2 3 1 5 1 2 1

x x x x P x x c x x x x P x x c x x x x P x x c      

Estimating P(x3 |e) for non-sampled node X3 :

X1 X7 X6 X5 X4 X2 X9 X8 X3

             ) , , | ( ) , , | ( ) , , | ( 3 1 ) | (

9 3 5 3 2 3 9 2 5 2 2 3 9 1 5 1 2 3 9 3

x x x x P x x x x P x x x x P x x P

CPCS54 Test Results

48

MSE vs. #samples (left) and time (right) Ergodic, |X|=54, D(Xi)=2, |C|=15, |E|=3 Exact Time = 30 sec using Cutset Conditioning

CPCS54, n=54, |C|=15, |E|=3 0.001 0.002 0.003 0.004 1000 2000 3000 4000 5000 # samples

Cutset Gibbs

CPCS54, n=54, |C|=15, |E|=3

0.0002 0.0004 0.0006 0.0008 5 10 15 20 25 Time(sec)

Cutset Gibbs

CPCS179 Test Results

49

MSE vs. #samples (left) and time (right) Non-Ergodic (1 deterministic CPT entry) |X| = 179, |C| = 8, 2<= D(Xi)<=4, |E| = 35 Exact Time = 122 sec using Cutset Conditioning

CPCS179, n=179, |C|=8, |E|=35

0.002 0.004 0.006 0.008 0.01 0.012 100 500 1000 2000 3000 4000 # samples Cutset Gibbs

CPCS179, n=179, |C|=8, |E|=35

0.002 0.004 0.006 0.008 0.01 0.012 20 40 60 80 Time(sec)

Cutset Gibbs

CPCS360b Test Results

50

MSE vs. #samples (left) and time (right) Ergodic, |X| = 360, D(Xi)=2, |C| = 21, |E| = 36 Exact Time > 60 min using Cutset Conditioning Exact Values obtained via Bucket Elimination

CPCS360b, n=360, |C|=21, |E|=36

0.00004 0.00008 0.00012 0.00016 200 400 600 800 1000 # samples

Cutset Gibbs

CPCS360b, n=360, |C|=21, |E|=36

0.00004 0.00008 0.00012 0.00016 1 2 3 5 10 20 30 40 50 60 Time(sec)

Cutset Gibbs

Random Networks

51

MSE vs. #samples (left) and time (right) |X| = 100, D(Xi) =2,|C| = 13, |E| = 15-20 Exact Time = 30 sec using Cutset Conditioning

RANDOM, n=100, |C|=13, |E|=15-20

0.0005 0.001 0.0015 0.002 0.0025 0.003 0.0035 200 400 600 800 1000 1200

# samples Cutset Gibbs

RANDOM, n=100, |C|=13, |E|=15-20

0.0002 0.0004 0.0006 0.0008 0.001 1 2 3 4 5 6 7 8 9 10 11 Time(sec) Cutset Gibbs

Coding Networks

Cutset Transforms Non-Ergodic Chain to Ergodic

52

MSE vs. time (right) Non-Ergodic, |X| = 100, D(Xi)=2, |C| = 13-16, |E| = 50 Sample Ergodic Subspace U={U1, U2,…Uk} Exact Time = 50 sec using Cutset Conditioning

x1 x2 x3 x4 u1 u2 u3 u4 p1 p2 p3 p4 y4 y3 y2 y1

Coding Networks, n=100, |C|=12-14

0.001 0.01 0.1 10 20 30 40 50 60 Time(sec)

IBP Gibbs Cutset

Non-Ergodic Hailfinder

53

MSE vs. #samples (left) and time (right) Non-Ergodic, |X| = 56, |C| = 5, 2 <=D(Xi) <=11, |E| = 0 Exact Time = 2 sec using Loop-Cutset Conditioning

HailFinder, n=56, |C|=5, |E|=1

0.0001 0.001 0.01 0.1 1 1 2 3 4 5 6 7 8 9 10

Time(sec)

Cutset Gibbs

HailFinder, n=56, |C|=5, |E|=1

0.0001 0.001 0.01 0.1 500 1000 1500

# samples Cutset Gibbs

CPCS360b - MSE

54

cpcs360b, N=360, |E|=[20-34], w*=20, MSE

0.000005 0.00001 0.000015 0.00002 0.000025

200 400 600 800 1000 1200 1400 1600 Time (sec)

Gibbs IBP |C|=26,fw=3 |C|=48,fw=2

MSE vs. Time Ergodic, |X| = 360, |C| = 26, D(Xi)=2 Exact Time = 50 min using BTE

Cutset Importance Sampling

• Apply Importance Sampling over cutset C

 

 

 

T t t T t t t

w T c Q e c P T e P

1 1

1 ) ( ) , ( 1 ) ( ˆ







T t t t i i

w c c T e c P

1

) , ( 1 ) | (  







T t t t i i

w e c x P T e x P

1

) , | ( 1 ) | ( 

where P(ct,e) is computed using Bucket Elimination

(Gogate & Dechter, 2005) and (Bidyuk & Dechter, 2006)

Likelihood Cutset Weighting (LCS)

• Z=Topological Order{C,E}
• Generating sample t+1:

56

For End If End ) ,..., | ( Else If : do For

1 1 1 1 1

1

    

    

t i t i t i i i t i i i

z z Z P z e ,z z z E Z Z Z

• computed while generating

sample t using bucket tree elimination

• can be memoized for some

number of instances K (based on memory available

KL[P(C|e), Q(C)] ≤ KL[P(X|e), Q(X)]

Pathfinder 1

57

Pathfinder 2

58

Link

59

Summary

• i.i.d. samples
• Unbiased estimator
• Generates samples fast
• Samples from Q
• Reject samples with

zero-weight

• Improves on cutset
• Dependent samples
• Biased estimator
• Generates samples

slower

• Samples from `P(X|e)
• Does not converge in

presence of constraints

• Improves on cutset

60

Importance Sampling Gibbs Sampling

CPCS360b

61

cpcs360b, N=360, |LC|=26, w*=21, |E|=15

1.E-05 1.E-04 1.E-03 1.E-02 2 4 6 8 10 12 14

Time (sec) MSE

LW AIS-BN Gibbs LCS IBP

LW – likelihood weighting LCS – likelihood weighting on a cutset

CPCS422b

62 1.0E-05 1.0E-04 1.0E-03 1.0E-02 10 20 30 40 50 60

MSE Time (sec)

cpcs422b, N=422, |LC|=47, w*=22, |E|=28

LW AIS-BN Gibbs LCS IBP

LW – likelihood weighting LCS – likelihood weighting on a cutset

Coding Networks

63

coding, N=200, P=3, |LC|=26, w*=21

1.0E-05 1.0E-04 1.0E-03 1.0E-02 1.0E-01 2 4 6 8 10

Time (sec) MSE

LW AIS-BN Gibbs LCS IBP

LW – likelihood weighting LCS – likelihood weighting on a cutset