Scribd
Upload
44 views

Forward Sampling: Sargur Srihari Srihari@buffalo - Edu

This document discusses forward sampling for probabilistic graphical models. It describes sampling from a Bayesian network by sampling nodes in topological order based on their conditional probability distributions. Samples can be used to estimate expectations and probabilities by taking averages. The number of samples required depends on the probability being estimated, with more samples needed for lower probability events. Rejection sampling can estimate conditional probabilities but is inefficient when the evidence has low probability.

Uploaded by

asdfasdffdsa
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
44 views

Forward Sampling: Sargur Srihari Srihari@buffalo - Edu

This document discusses forward sampling for probabilistic graphical models. It describes sampling from a Bayesian network by sampling nodes in topological order based on their conditional probability distributions. Samples can be used to estimate expectations and probabilities by taking averages. The number of samples required depends on the probability being estimated, with more samples needed for lower probability events. Rejection sampling can estimate conditional probabilities but is inefficient when the evidence has low probability.

Uploaded by

asdfasdffdsa
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 13

Probabilistic Graphical Models Srihari

Forward Sampling

Sargur Srihari
[email protected]

1
Probabilistic Graphical Models Srihari

Topics

•  Forward Sampling
•  Sampling from a Bayesian Network
•  Sampling from a Discrete Distribution
•  Use of samples
•  Analysis of Error
•  Conditional Probability Queries

2
Probabilistic Graphical Models Srihari

Forward Sampling: Plan of Discussion


•  Forward sampling
–  It is the simplest particle generation approach
–  We generate samples ξ[1],..,ξ[M] from P(χ)
ξ denotes a full assignment to variables in χ, i.e., χ εVal (χ)

•  Plan of discussion
–  Generating particles from PB (χ) by sampling a BN
–  No of particles needed to get a good approximation
of the expectation of a target function f
–  Difficulties in generating samples from posterior
PB (χ|e)
•  In undirected models even generating a sample from the
prior distribution is a difficult task
Probabilistic Graphical Models Srihari

Sampling from a Bayesian Network


•  A very simple process
•  Sample nodes in some order so that by the
time we sample a node we have values for all
of its parents
•  We can then sample from the distribution
specified by the CPD
•  Need ability to sample from the distributions
underlying the CPD
–  Straightforward for discrete case
–  Subtler for continuous measures 4
Probabilistic Graphical Models Srihari

Ex: Forward (Ancestral) Sampling

1 Sample D {0.6 vs 0.4}àd0


2 Sample I {0.7 vs 0.3}ài0
3 Sample G {0.3, 0.4, 0.3}àg2
etc, etc

5
Probabilistic Graphical Models Srihari

Sampling from a Discrete Distribution


•  Split [0,1] interval into bins whose sizes are
determined by the probabilities P(xi), i=1,..,k
–  Partition the interval into k subintervals
•  Generate a sample s uniformly from the interval
•  If s is in the ith interval then sampled value is xi
•  Example: sample from
P(x={1,2,3,4})={0.3,0.1,0.4,0.2}

6
Probabilistic Graphical Models Srihari

Using samples to find expectations


•  To compute an expectation E[f (x)] = ∑ p(x)f (x)
x

–  Using basic convergence bounds, we know that


from a set of particles D ={ξ[1]),..,(ξ[M]} generated via
sampling we can estimate it as
M
1
Ê D (f ) =
M
∑ f (ξ[m])
m=1

•  In the case where our task is to compute P(y)


•  for event Y=y relative to P for some Y⊆X and y ∈ Val(Y)
–  this estimate is simply the fraction of particles where
we have seen the event y
1 M y[m] denotes ξ[m]Y
P̂D (y) =
M
∑ I {y[m] = y} i.e., assignment to Y in the particle ξ[m]
m =1
Probabilistic Graphical Models Srihari

Analysis of Error
•  Quality of the estimate depends heavily on the
no. of particles generated
•  How many particles are needed to obtain a
certain performance guarantee?
•  Focus analysis on the case where we need to
estimate P(y)

8
Probabilistic Graphical Models Srihari

How many samples are required?


For M ind Bernoulli trials with prob (success)=p
SD=X[1]+..X[M], TD=(1/M)SD
1.  From Hoeffding bound P(TD>p+ε)≤exp(-2Mε2)
P(TD<p - ε)≤exp(-2Mε2)

–  Error is bounded by ε with probability of at least 1-δ


–  The estimator with (ε,δ) reliability is ln(2 / δ)
M ≥
2ε 2
2.  From Chernoff bound
–  For a given ε, no of samples needed to guarantee a
certain error probability δ is
ln(2 / δ)
M ≥3
P(y)ε 2

•  M grows inversely with probability P(y)


–  Thus, cannot determine no. of samples needed
Probabilistic Graphical Models Srihari

Conditional Probability Queries


•  So far we have discussed the problem of
estimating marginal probabilities
–  i.e., Probability of event Y=y relative to the original
joint distribution
•  In general we are interested in conditional
distributions of the form p(y|E=e)
•  This estimation task is significantly harder
•  One approach is called Rejection sampling
10
Probabilistic Graphical Models Srihari

Rejection Sampling
•  Generate samples from posterior P(χ|e)
–  Can do this by generating samples x from P(X)
–  Reject any sample that is not compatible with e
–  Resulting samples are from posterior P(χ|e)
•  Problem: no of unrejected particles is small
–  Expected no of unrejected from original M is MP(e)
•  Ex: if P(e)=0.001 & M=10,000, only 10 unrejected samples
–  Conversely, to obtain at least M* unrejected
particles, we need to generate on an average
M=M*/P(e) samples from P(X)
11
Probabilistic Graphical Models Srihari

Low probability evidence is common


•  In many applications, low probability evidence
is the rule rather than the exception
•  Example of medical diagnosis
–  Any set of symptoms typically has a low probability
•  As no of variables k increases, probability of evidence k
decreases exponentially with k

12
Probabilistic Graphical Models Srihari

Indirect Approach also has limitation


•  Alternative approach to determining P(y|E=e) is
to use a separate estimator for P(e) and for
P(y,e) and then compute the ratio
–  If both quantities gave low relative error then their
ratio also will have low relative error
•  No. of samples required to get a low relative
error also grows linearly with 1/P(e)

13

You might also like