Lecture Note
Lecture 16: Inference in Bayesian Networks
-
University:
Massachusetts Institute of Technology -
Course:
6.825 | Techniques in Artificial Intelligence (SMA 5504) Academic year:
2019
-
Views:
421
Pages:
74
Author:
Jessie Yates
6.825 Techniques in Artificial Intelligence
Inference in Bayesian Networks
Lecture 16 • 1
Now that we know what the semantics of Bayes nets are; what it means when we
have one, we need to understand how to use it. Typically, we’ll be in a situation
in which we have some evidence, that is, some of the variables are instantiated,
and we want to infer something about the probability distribution of some other
variables.
1 6.825 Techniques in Artificial Intelligence
Inference in Bayesian Networks
• Exact inference
Lecture 16 • 2
In exact inference, we analytically compute the conditional probability distribution
over the variables of interest.
2 6.825 Techniques in Artificial Intelligence
Inference in Bayesian Networks
• Exact inference
• Approximate inference
Lecture 16 • 3
But sometimes, that’s too hard to do, in which case we can use approximation
techniques based on statistical sampling.
3 Query Types
Given a Bayesian network, what questions might we
want to ask?
Lecture 16 • 4
Given a Bayesian network, what kinds of questions might we want to ask?
4 Query Types
Given a Bayesian network, what questions might we
want to ask?
• Conditional probability query: P(x | e)
Lecture 16 • 5
The most usual is a conditional probability query. Given instantiations for some of
the variables (we’ll use e here to stand for the values of all the instantiated
variables; it doesn’t have to be just one), what is the probability that node X has
a particular value x?
5 Query Types
Given a Bayesian network, what questions might we
want to ask?
• Conditional probability query: P(x | e)
• Maximum a posteriori probability:
What value of x maximizes P(x|e) ?
Lecture 16 • 6
Another interesting question you might ask is, what is the most likely explanation
for some evidence? We can think of that as the value of node X (or of some
group of nodes) that maximizes the probability that you would have seen the
evidence you did. This is called the maximum a posteriori probability or MAP
query.
6 Query Types
Given a Bayesian network, what questions might we
want to ask?
• Conditional probability query: P(x | e)
• Maximum a posteriori probability:
What value of x maximizes P(x|e) ?
General question: What’s the whole probability
distribution over variable X given evidence e,
P(X | e)?
Lecture 16 • 7
In our discrete probability situation, the only way to answer a MAP query is to
compute the probability of x given e for all possible values of x and see which
one is greatest.
7 Query Types
Given a Bayesian network, what questions might we
want to ask?
• Conditional probability query: P(x | e)
• Maximum a posteriori probability:
What value of x maximizes P(x|e) ?
General question: What’s the whole probability
distribution over variable X given evidence e,
P(X | e)?
Lecture 16 • 8
So, in general, we’d like to be able to compute a whole probability distribution over
some variable or variables X, given instantiations of a set of variables e.
8 Using the joint distribution
To answer any query involving a conjunction of
variables, sum over the variables not involved in
the query.
Lecture 16 • 9
Given the joint distribution over the variables, we can easily answer any question
about the value of a single variable by summing (or marginalizing) over the
other variables.
9 Using the joint distribution
To answer any query involving a conjunction of
variables, sum over the variables not involved in
the query.
Pr(d) = ∑ Pr(a,b,c,d )
ABC
=
∑
∑
∑ Pr(A = a ∧ B = b ∧C = c)
a∈dom(A ) b∈dom(B ) c∈dom(C )
Lecture 16 • 10
So, in a domain with four variables, A, B, C, and D, the probability that variable D
has value d is the sum over all possible combinations of values of the other
three variables of the joint probability of all four values. This is exactly the
same as the procedure we went through in the last lecture, where to compute the
probability of cavity, we added up the probability of cavity and toothache and
the probability of cavity and not toothache.
10 Using the joint distribution
To answer any query involving a conjunction of
variables, sum over the variables not involved in
the query.
Pr(d) = ∑ Pr(a,b,c,d )
ABC
=
∑
∑
∑ Pr(A = a ∧ B = b ∧C = c)
a∈dom(A ) b∈dom(B ) c∈dom(C )
Lecture 16 • 11
In general, we’ll use the first notation, with a single summation indexed by a list of
variable names, and a joint probability expression that mentions values of those
variables. But here we can see the completely written-out definition, just so we
all know what the shorthand is supposed to mean.
11 Using the joint distribution
To answer any query involving a conjunction of
variables, sum over the variables not involved in
the query.
Pr(d) = ∑ Pr(a,b,c,d )
ABC
=
∑
∑
∑ Pr(A = a ∧ B = b ∧C = c)
a∈dom(A ) b∈dom(B ) c∈dom(C )
∑ Pr(a,b,c,d )
Pr(b,d) AC
Pr(d | b) =
=
Pr(b)
∑ Pr(a,b,c,d)
ACD
Lecture 16 • 12
To compute a conditional probability, we reduce it to a ratio of conjunctive queries
using the definition of conditional probability, and then answer each of those
queries by marginalizing out the variables not mentioned.
12 Using the joint distribution
To answer any query involving a conjunction of
variables, sum over the variables not involved in
the query.
Pr(d) = ∑ Pr(a,b,c,d )
ABC
=
∑
∑
∑ Pr(A = a ∧ B = b ∧C = c)
a∈dom(A ) b∈dom(B ) c∈dom(C )
∑ Pr(a,b,c,d )
Pr(b,d) AC
Pr(d | b) =
=
Pr(b)
∑ Pr(a,b,c,d)
ACD
Lecture 16 • 13
In the numerator, here, you can see that we’re only summing over variables A and
C, because b and d are instantiated in the query.
13 Simple Case
Lecture 16 • 14
We’re going to learn a general purpose algorithm for answering these joint queries
fairly efficiently. We’ll start by looking at a very simple case to build up our
intuitions, then we’ll write down the algorithm, then we’ll apply it to a more
complex case.
14 Simple Case
A
B
C
D
Lecture 16 • 15
Okay. Here’s our very simple case. It’s a bayes net with four nodes, arranged in a
chain.
15 Simple Case
A
Pr( d ) =
B
C
D
∑ Pr(a, b, c, d )
ABC
Lecture 16 • 16
So, we know from before that the probability that variable D has some value little d
is the sum over A, B, and C of the joint distribution, with d fixed.
16 Simple Case
A
B
C
D
Pr(d ) = ∑Pr(a, b, c, d )
ABC
= ∑Pr(d | c) Pr(c | b) Pr(b | a) Pr(a)
ABC
Lecture 16 • 17
Now, using the chain rule of Bayesian networks, we can write down the joint
probability as a product over the nodes of the probability of each node’s value
given the values of its parents. So, in this case, we get P(d|c) times P(c|b) times
P(b|a) times P(a).
17 Simple Case
A
B
C
D
Pr(d ) = ∑Pr(a, b, c, d )
ABC
= ∑Pr(d | c) Pr(c | b) Pr(b | a) Pr(a)
ABC
Lecture 16 • 18
This expression gives us a method for answering the query, given the conditional
probabilities that are stored in the net. And this method can be applied directly
to any other bayes net. But there’s a problem with it: it requires enumerating
all possible combinations of assignments to A, B, and C, and then, for each one,
multiplying the factors for each node. That’s an enormous amount of work and
we’d like to avoid it if at all possible.
18 Simple Case
A
B
C
D
Pr(d ) = ∑Pr(a, b, c, d )
ABC
= ∑Pr(d | c) Pr(c | b) Pr(b | a) Pr(a)
ABC
= ∑∑∑Pr(d | c) Pr(c | b) Pr(b | a) Pr(a)
C
B
A
Lecture 16 • 19
So, we’ll try rewriting the expression into something that might be more efficient to
evaluate. First, we can make our summation into three separate summations,
one over each variable.
19 Simple Case
A
B
Pr(d) =
C
D
∑ Pr(a,b,c,d)
ABC
=
∑ Pr(d | c)Pr(c | b)Pr(b | a)Pr(a)
ABC
= ∑ ∑ ∑ Pr(d | c)Pr(c | b)Pr(b | a)Pr(a)
C
B
A
= ∑ Pr(d | c)∑ Pr(c | b)∑ Pr(b | a)Pr(a)
C
B
A
Lecture 16 • 20
Then, by distributivity of addition over multiplication, we can push the summations
in, so that the sum over A includes all the terms that mention A, but no others,
and so on. It’s pretty clear that this expression is the same as the previous one in
value, but it can be evaluated more efficiently. We’re still, eventually,
enumerating all assignments to the three variables, but we’re doing somewhat
fewer multiplications than before. So this is still not completely satisfactory.
20 Simple Case
A
B
C
D
Pr(d) = ∑ Pr(d | c)∑ Pr(c | b)∑ Pr(b | a)Pr(a)
C
B
A
Pr(b1 | a1 ) Pr(a1 ) Pr(b1 | a 2 ) Pr(a2 )
Pr(b2 | a1 ) Pr(a1 ) Pr(b2 | a2 ) Pr(a2 )
Lecture 16 • 21
If you look, for a minute, at the terms inside the summation over A, you’ll see that
we’re doing these multiplications over for each value of C, which isn’t
necessary, because they’re independent of C. Our idea, here, is to do the
multiplications once and store them for later use. So, first, for each value of A
and B, we can compute the product, generating a two dimensional matrix.
21 Simple Case
A
B
C
D
Pr(d) = ∑ Pr(d | c)∑ Pr(c | b)∑ Pr(b | a)Pr(a)
C
B
A
∑ Pr(b1 | a) Pr(a)
A
Pr(b | a) Pr(a)
2
∑
A
Lecture 16 • 22
Then, we can sum over the rows of the matrix, yielding one value of the sum for
each possible value of b.
22 Simple Case
A
B
C
D
Pr(d) = ∑ Pr(d | c)∑ Pr(c | b)∑ Pr(b | a)Pr(a)
C
B
A
f1 (b)
Lecture 16 • 23
We’ll call this set of values, which depends on b, f1 of b.
23 Simple Case
B
C
D
Pr(d) = ∑ Pr(d | c)∑ Pr(c | b) f1 (b)
C
B
f2 (c)
Lecture 16 • 24
Now, we can substitute f1 of b in for the sum over A in our previous expression.
And, effectively, we can remove node A from our diagram. Now, we express
the contribution of b, which takes the contribution of a into account, as f_1 of b.
24 Simple Case
B
C
D
Pr(d) = ∑ Pr(d | c)∑ Pr(c | b) f1 (b)
C
B
f2 (c)
Lecture 16 • 25
We can continue the process in basically the same way. We can look at the
summation over b and see that the only other variable it involves is c. We can
summarize those products as a set of factors, one for each value of c. We’ll call
those factors f_2 of c.
25 Simple Case
C
D
Pr(d) = ∑ Pr(d | c) f2 (c)
C
Lecture 16 • 26
We substitute f_2 of c into the formula, remove node b from the diagram, and now
we’re down to a simple expression in which d is known and we have to sum
over values of c.
26 Variable Elimination Algorithm
Given a Bayesian network, and an elimination order
for the non-query variables
Lecture 16 • 27
That was a simple special case. Now we can look at the algorithm in the general
case. Let’s assume that we’re given a Bayesian network and an ordering on the
variables that aren’t fixed in the query. We’ll come back later to the question of
the influence of the order, and how we might find a good one.
27 Variable Elimination Algorithm
Given a Bayesian network, and an elimination order
for the non-query variables, compute
∑∑K ∑ ∏ Pr(x
X1 X 2
Xm
j
| Pa(x j ))
j
Lecture 16 • 28
We can express the probability of the query variables as a sum over each value of
each of the non-query variables of a product over each node in the network, of
the probability that that variable has the given value given the values of its
parents.
28 Variable Elimination Algorithm
Given a Bayesian network, and an elimination order
for the non-query variables, compute
∑∑K ∑ ∏ Pr(x
X1 X 2
Xm
j
| Pa(x j ))
j
For i = m downto 1
Lecture 16 • 29
So, we’ll eliminate the variables from the inside out. Starting with variable Xm and
finishing with variable X1.
29 Variable Elimination Algorithm
Given a Bayesian network, and an elimination order
for the non-query variables, compute
∑∑K ∑ ∏ Pr(x
X1 X 2
Xm
j
| Pa(x j ))
j
For i = m downto 1
• remove all the factors that mention Xi
Lecture 16 • 30
To eliminate variable Xi, we start by gathering up all of the factors that mention Xi,
and removing them from our set of factors. Let’s say there are k such factors.
30 Variable Elimination Algorithm
Given a Bayesian network, and an elimination order
for the non-query variables, compute
∑∑K ∑ ∏ Pr(x
X1 X 2
Xm
j
| Pa(x j ))
j
For i = m downto 1
• remove all the factors that mention Xi
• multiply those factors, getting a value for each
combination of mentioned variables
Lecture 16 • 31
Now, we make a k+1 dimensional table, indexed by Xi as well as each of the other
variables that is mentioned in our set of factors.
31 Variable Elimination Algorithm
Given a Bayesian network, and an elimination order
for the non-query variables, compute
∑∑K ∑ ∏ Pr(x
X1 X 2
Xm
j
| Pa(x j ))
j
For i = m downto 1
• remove all the factors that mention Xi
• multiply those factors, getting a value for each
combination of mentioned variables
• sum over Xi
Lecture 16 • 32
We then sum the table over the Xi dimension, resulting in a k-dimensional table.
32 Variable Elimination Algorithm
Given a Bayesian network, and an elimination order
for the non-query variables, compute
∑∑K ∑ ∏ Pr(x
X1 X 2
Xm
j
| Pa(x j ))
j
For i = m downto 1
• remove all the factors that mention Xi
• multiply those factors, getting a value for each
combination of mentioned variables
• sum over Xi
• put this new factor into the factor set
Lecture 16 • 33
This table is our new factor, and we put a term for it back into our set of factors.
33 Variable Elimination Algorithm
Given a Bayesian network, and an elimination order
for the non-query variables, compute
∑∑K ∑ ∏ Pr(x
X1 X 2
Xm
j
| Pa(x j ))
j
For i = m downto 1
• remove all the factors that mention Xi
• multiply those factors, getting a value for each
combination of mentioned variables
• sum over Xi
• put this new factor into the factor set
Lecture 16 • 34
Once we’ve eliminated all the summations, we have the desired value.
34 One more example
risky
Smoke
Visit
Lung
cancer
Tuber
culosis
Bronch
chest
Abnrm
Xray
itis
Dysp
nea
Lecture 16 • 35
Here’s a more complicated example, to illustrate the variable elimination algorithm
in a more general case. We have this big network that encodes a domain for
diagnosing lung disease. (Dyspnea, as I understand it, is shortness of breath).
35 One more example
risky
Smoke
Visit
Lung
cancer
Tuber
culosis
Bronch
chest
Abnrm
Dysp
Xray
Pr(d) =
∑
A,B,L,T,S ,X,V
itis
nea
Pr(d | a,b) Pr(a | t,l) Pr(b | s) Pr(l | s) Pr(s)
Pr(x | a) Pr(t | v) Pr(v)
Lecture 16 • 36
We’ll do variable elimination on this graph using elimination order A, B, L, T, S, X,
V.
36 One more example
risky
Smoke
Visit
Lung
cancer
Tuber
culosis
Bronch
chest
Abnrm
itis
Dysp
Xray
Pr(d) =
∑
A,B,L,T,S ,X
nea
Pr(d | a,b) Pr(a | t,l) Pr(b | s) Pr(l | s) Pr(s)
Pr(x | a)∑ Pr(t | v) Pr(v)
V
f1 (t)
Lecture 16 • 37
So, we start by eliminating V. We gather the two terms that mention V and see that
they also involve variable T. So, we compute the product for each value of T,
and summarize those in the factor f1 of T.
37 One more example
Smoke
Lung
cancer
Tuber
culosis
Bronch
chest
Abnrm
Dysp
Xray
Pr(d) =
∑
A,B,L,T,S ,X
itis
nea
Pr(d | a,b) Pr(a | t,l) Pr(b | s) Pr(l | s) Pr(s)
Pr(x | a) f1 (t)
Lecture 16 • 38
Now we can substitute that factor in for the summation, and remove the node from
the network.
38 One more example
Smoke
Lung
cancer
Tuber
culosis
Bronch
chest
Abnrm
Dysp
Xray
Pr(d) =
itis
nea
Pr(d | a,b) Pr(a | t,l) Pr(b | s) Pr(l | s) Pr(s) f1 (t)
∑ ∑ Pr(x | a)
A,B,L,T,S
X
1
Lecture 16 • 39
The next variable to be eliminated is X. There is actually only one term involving
X, and it also involves variable A. So, for each value of A, we compute the sum
over X of P(x|a). But wait! We know what this value is! If we fix a and sum
over x, these probabilities have to add up to 1.
39 One more example
Smoke
Lung
cancer
Tuber
culosis
Bronch
chest
Abnrm
itis
Dysp
nea
Pr(d) =
∑ Pr(d | a,b) Pr(a | t,l) Pr(b | s) Pr(l | s) Pr(s) f (t)
1
A,B,L,T,S
Lecture 16 • 40
So, rather than adding another factor to our expression, we can just remove the
whole sum. In general, the only nodes that will have an influence on the
probability of D are its ancestors.
40 One more example
Smoke
Lung
cancer
Tuber
culosis
Bronch
chest
Abnrm
itis
Dysp
nea
Pr(d) =
∑
A,B,L,T
Pr(d | a,b) Pr(a | t,l) f1 (t)∑ Pr(b | s) Pr(l | s) Pr(s)
S
Lecture 16 • 41
Now, it’s time to eliminate S. We find that there are three terms involving S, and
we gather them into the sum. These three terms involve two other variables, B
and L. So we have to make a factor that specifies, for each value of B and L, the
value of the sum of products.
41 One more example
Smoke
Lung
cancer
Tuber
culosis
Bronch
chest
Abnrm
itis
Dysp
nea
Pr(d) =
∑
Pr(d | a,b) Pr(a | t,l) f1 (t)∑ Pr(b | s) Pr(l | s) Pr(s)
A,B,L,T
S
f2 (b,l)
Lecture 16 • 42
We’ll call that factor f_2 of b and l.
42 One more example
Lung
cancer
Tuber
culosis
Bronch
chest
Abnrm
itis
Dysp
nea
Pr(d) =
∑ Pr(d | a,b) Pr(a | t,l) f (t) f (b,l)
1
2
A,B,L,T
Lecture 16 • 43
Now we can substitute that factor back into our expression. We can also eliminate
node S. But in eliminating S, we’ve added a direct dependency between L and
B (they used to be dependent via S, but now the dependency is encoded
explicitly in f2(b). We’ll show that in the graph by drawing a line between the
two nodes. It’s not exactly a standard directed conditional dependence, but it’s
still useful to show that they’re coupled.
43 One more example
Lung
cancer
Tuber
culosis
Bronch
chest
Abnrm
itis
Dysp
nea
Pr(d) =
∑ Pr(d | a,b) f (b,l)∑ Pr(a | t,l) f (t)
2
A,B,L
1
T
f3 (a,l)
Lecture 16 • 44
Now we eliminate T. It involves two terms, which themselves involve variables A
and L. So we make a new factor f3 of A and L.
44 One more example
Lung
cancer
Bronch
chest
Abnrm
itis
Dysp
nea
Pr(d) =
∑ Pr(d | a,b) f (b,l) f (a,l)
2
3
A,B,L
Lecture 16 • 45
We can substitute in that factor and eliminate T. We’re getting close!
45 One more example
Lung
cancer
Bronch
chest
Abnrm
itis
Dysp
nea
Pr(d) = ∑ Pr(d | a,b)∑ f2 (b,l) f3 (a,l)
A ,B
L
f4 (a,b)
Lecture 16 • 46
Next we eliminate L. It involves these two factors, which depend on variables A
and B. So we make a new factor, f4 of A and B, and substitute it in. We
remove node L, but couple A and B.
46 One more example
Bronch
chest
Abnrm
itis
Dysp
nea
Pr(d) = ∑ Pr(d | a,b) f4 (a,b)
A ,B
Lecture 16 • 47
At this point, we could just do the summations over A and B and be done. But to
finish out the algorithm the way a computer would, it’s time to eliminate variable B.
47 One more example
Bronch
chest
Abnrm
itis
Dysp
nea
Pr(d) = ∑ ∑ Pr(d | a,b) f4 (a,b)
A
B
f 5 (a)
Lecture 16 • 48
It involves both of our remaining terms, and it seems to depend on variables A and
D. However, in this case, we’re interested in the probability of a particular
value, little d of D, and so the variable d is instantiated. Thus, we can treat it as
a constant in this expression, and we only need to generate a factor over a,
which we’ll call f5 of a. And we can now, in some sense, remove D from our
network as well (because we’ve already factored it into our answer).
48 One more example
chest
Abnrm
Pr(d) = ∑ f 5 (a)
A
Lecture 16 • 49
Finally, to get the probability that variable D has value little d, we simply sum
factor f5 over all values of a. Yay! We did it.
49 Properties of Variable Elimination
Lecture 16 • 50
Let’s see how the variable elimination algorithm performs, both in theory and in
practice.
50 Properties of Variable Elimination
• Time is exponential in size of largest factor
Lecture 16 • 51
First of all, it’s pretty easy to see that it runs in time exponential in the number of
variables involved in the largest factor. Creating a factor with k variables
involves making a k+1 dimensional table. If you have b values per variable,
that’s a table of size b^(k+1). To make each entry, you have to multiply at most
n numbers, where n is the number of nodes. We have to do this for each
variable to be eliminated (which is usually close to n). So we have something
like time = O(n^2 b^k).
51 Properties of Variable Elimination
• Time is exponential in size of largest factor
• Bad elimination order can generate huge factors
Lecture 16 • 52
How big the factors are depends on the elimination order. You’ll see in one of the
recitation exercises just how dramatic the difference in factor sizes can be. A
bad elimination order can generate huge factors.
52 Properties of Variable Elimination
• Time is exponential in size of largest factor
• Bad elimination order can generate huge factors
• NP Hard to find the best elimination order
• Even the best elimination order may generate large
factors
Lecture 16 • 53
So, we’d like to use the elimination order that generates the smallest factors.
Unfortunately, it turns out to be NP hard to find the best elimination order.
53 Properties of Variable Elimination
• Time is exponential in size of largest factor
• Bad elimination order can generate huge factors
• NP Hard to find the best elimination order
• Even the best elimination order may generate large
factors
• There are reasonable heuristics for picking an
elimination order (such as choosing the variable
that results in the smallest next factor)
Lecture 16 • 54
At least, there are some fairly reasonable heuristics for choosing an elimination
order. It’s usually done dynamically. So, rather than fixing the elimination
order in advance, as we suggested in the algorithm description, you can pick the
next variable to be eliminated depending on the situation. In particular, one
reasonable heuristic is to pick the variable to eliminate next that will result in
the smallest factor. This greedy approach won’t always be optimal, but it’s not
usually too bad.
54 Properties of Variable Elimination
• Time is exponential in size of largest factor
• Bad elimination order can generate huge factors
• NP Hard to find the best elimination order
• Even the best elimination order may generate large
factors
• There are reasonable heuristics for picking an
elimination order (such as choosing the variable
that results in the smallest next factor)
• Inference in polytrees (nets with no cycles) is linear
in size of the network (the largest CPT)
Lecture 16 • 55
There is one case where Bayes net inference in general, and the variable elimination
algorithm in particular is fairly efficient, and that’s when the network is a
polytree. A polytree is a network with no cycles. That is, a network in which,
for any two nodes, there is only one path between them. In a polytree, inference
is linear in the size of the network, where the size of the network is defined to be
the size of the largest conditional probability table (or exponential in the
maximum number of parents of any node). In a polytree, the optimal
elimination order is to start at the root nodes, and work downwards, always
eliminating a variable that no longer has any parents. In doing so, we never
introduce additional connections into the network.
55 Properties of Variable Elimination
• Time is exponential in size of largest factor
• Bad elimination order can generate huge factors
• NP Hard to find the best elimination order
• Even the best elimination order may generate large
factors
• There are reasonable heuristics for picking an
elimination order (such as choosing the variable
that results in the smallest next factor)
• Inference in polytrees (nets with no cycles) is linear
in size of the network (the largest CPT)
• Many problems with very large nets have only small
factors, and thus efficient inference
Lecture 16 • 56
So, inference in polytrees is efficient, and even in many large non-polytree
networks, it’s possible to keep the factors small, and therefore to do inference
relatively efficiently.
56 Properties of Variable Elimination
• Time is exponential in size of largest factor
• Bad elimination order can generate huge factors
• NP Hard to find the best elimination order
• Even the best elimination order may generate large
factors
• There are reasonable heuristics for picking an
elimination order (such as choosing the variable
that results in the smallest next factor)
• Inference in polytrees (nets with no cycles) is linear
in size of the network (the largest CPT)
• Many problems with very large nets have only small
factors, and thus efficient inference
Lecture 16 • 57
When the network is such that the factors are, of necessity, large, we’ll have to turn
to a different class of methods.
57 Sampling
Lecture 16 • 58
Another strategy, which is a theme that comes up also more and more in AI
actually, is to say, well, we didn't really want the right answer anyway. Let's try
to do an approximation. And you can also show that it's computationally hard
to get an approximation that's within epsilon of the answer that you want, but
again that doesn't keep us from trying.
58 Sampling
To get approximate answer we can do stochastic simulation
(sampling).
A
B
P(A) = 0.4
C
D
Lecture 16 • 59
So, the other thing that we can do is the stochastic simulation or sampling. In
sampling, what we do is we look at the root nodes of our graph, and attached to
this root node is some probability that A is going to be true, right? Maybe it's
.4. So we flip a coin that comes up heads with probability .4 and see if we get
true or false.
59 Sampling
To get approximate answer we can do stochastic simulation
(sampling).
A
B
P(A) = 0.4
C
D
A
B
C
D
T
F
T
T
…
•Flip a coin where P(T)=0.4, assume
we get T, use that value for A
Lecture 16 • 60
We flip our coin, let's say, and we get true for A -- this time. And now, given the
assignment of true to A, we look in the conditional probability table for B given
A = true, and that gives us a probability for B.
60 Sampling
To get approximate answer we can do stochastic simulation
(sampling).
A
B
P(A) = 0.4
C
D
A
B
C
D
T
F
T
T
…
•Flip a coin where P(T)=0.4, assume
we get T, use that value for A
•Given A=T, lookup P(B|A=T) and flip
a coin with that prob., assume we get
F
Lecture 16 • 61
Now, we flip a coin with that probability. Say we get False. We enter that into the
table.
61 Sampling
To get approximate answer we can do stochastic simulation
(sampling).
A
B
P(A) = 0.4
C
D
A
B
C
D
T
F
T
T
…
•Flip a coin where P(T)=0.4, assume
we get T, use that value for A
•Given A=T, lookup P(B|A=T) and flip
a coin with that prob., assume we get
F
•Similarly for C and D
Lecture 16 • 62
We do the same thing for C, and let’s say we get True.
62 Sampling
To get approximate answer we can do stochastic simulation
(sampling).
A
B
P(A) = 0.4
C
D
A
B
C
D
T
F
T
T
…
•Flip a coin where P(T)=0.4, assume
we get T, use that value for A
•Given A=T, lookup P(B|A=T) and flip
a coin with that prob., assume we get
F
•Similarly for C and D
Lecture 16 • 63
Now, we look in the CPT for D given B and C, for the case where B is false and C
is true, and we flip a coin with that probability, in order to get a value for D.
63 Sampling
To get approximate answer we can do stochastic simulation
(sampling).
A
B
P(A) = 0.4
C
D
A
B
C
D
T
F
T
T
…
•Flip a coin where P(T)=0.4, assume
we get T, use that value for A
•Given A=T, lookup P(B|A=T) and flip
a coin with that prob., assume we get
F
•Similarly for C and D
•We get one sample from joint
distribution of these four vars
Lecture 16 • 64
So, there's one sample from the joint distribution of these four variables. And you
can just keep doing this, all day and all night, and generate a big pile of
samples, using that algorithm. And now you can ask various questions.
64 Sampling
To get approximate answer we can do stochastic simulation
(sampling).
A
B
P(A) = 0.4
C
D
Estimate:
P*(D|A) = #D,A / #A
A
B
C
D
T
F
T
T
…
•Flip a coin where P(T)=0.4, assume
we get T, use that value for A
•Given A=T, lookup P(B|A=T) and flip
a coin with that prob., assume we get
F
•Similarly for C and D
•We get one sample from joint
distribution of these four vars
Lecture 16 • 65
Let's say you want to know the probability of D given A. How would you answer - given all the examples -- what would you do to compute the probability of D
given A? You would just count. You’d count the number of cases in which A
and D were true, and you’d divide that by the number of cases in which A was
true, and that would give you an unbiased estimate of the probability of D given
A. The more samples, the more confidence you’d have that the estimated
probability is close to the true one.
65 Estimation
• Some probabilities are easier than others to
estimate
Lecture 16 • 66
It's going to turn out that some probabilities are easier than other ones to estimate.
66 Estimation
• Some probabilities are easier than others to
estimate
• In generating the table, the rare events will not be
well represented
Lecture 16 • 67
Exactly because of the process we’re using to generate the samples, the majority of
them will be the typical cases. Oh, it's someone with a cold, someone with a
cold, someone with a cold, someone with a cold, someone with a cold, someone
with malaria, someone with a cold, someone with a cold. So the rare results
are not going to come up very often. And so doing this sampling naively can
make it really hard to estimate the probability of a rare event. If it's something
that happens one in ten thousand times, well, you know for sure you're going to
need, some number of tens of thousands of samples to get even a reasonable
estimate of that probability.
67 Estimation
• Some probabilities are easier than others to
estimate
• In generating the table, the rare events will not be
well represented
• P(Disease| spots-on-your-tongue, sore toe)
Lecture 16 • 68
Imagine that you want to estimate the probability of some disease given -- oh, I
don't know -- spots on your tongue and a sore toe. Somebody walks in and they
have a really peculiar set of symptoms, and you want to know what's the
probability that they have some disease.
68 Estimation
• Some probabilities are easier than others to
estimate
• In generating the table, the rare events will not be
well represented
• P(Disease| spots-on-your-tongue, sore toe)
• If spots-on-your-tongue and sore toe are not root
nodes, you would generate a huge table but the
cases of interest would be very sparse in the table
Lecture 16 • 69
Well, if the symptoms are root nodes, it's easy. If the symptoms were root nodes,
you could just assign the root nodes to have their observed values and then
simulate the rest of the network as before.
69 Estimation
• Some probabilities are easier than others to
estimate
• In generating the table, the rare events will not be
well represented
• P(Disease| spots-on-your-tongue, sore toe)
• If spots-on-your-tongue and sore toe are not root
nodes, you would generate a huge table but the
cases of interest would be very sparse in the table
Lecture 16 • 70
But if the symptoms aren't root nodes then if you do naïve sampling, you would
generate a giant table of samples, and you'd have to go and look and say, gosh,
how many cases do I have where somebody has spots on their tongue and a sore
toe; and the answer would be, well, maybe zero or not very many.
70 Estimation
• Some probabilities are easier than others to
estimate
• In generating the table, the rare events will not be
well represented
• P(Disease| spots-on-your-tongue, sore toe)
• If spots-on-your-tongue and sore toe are not root
nodes, you would generate a huge table but the
cases of interest would be very sparse in the table
• Importance sampling lets you focus on the set of
cases that are important to answering your
question
Lecture 16 • 71
There’s a technique called importance sampling, which allows you to draw
examples from a distribution that’s going to be more helpful and then reweight
them so that you can still get an unbiased estimate of the desired conditional
probability. It’s a bit beyond the scope of this class to get into the details, but
it’s an important and effective idea.
71 Recitation Problem
• Do the variable elimination algorithm on the net
below using the elimination order A,B,C (that is,
eliminate node C first). In computing P(D=d), what
factors do you get?
• What if you wanted to compute the whole marginal
distribution P(D)?
A
B
C
D
Lecture 16 • 72
Here’s the network we started with. We used elimination order C, B, A (we
eliminated A first). Now we’re going to explore what happens when we
eliminate the variables in the opposite order. First, work on the case we did,
where we’re trying to calculate the probability that node D takes on a particular
value, little d. Remember that little d is a constant in this case. Now, do the
case where we’re trying to find the whole distribution over D, so we don’t know
a particular value for little d.
72 Another Recitation Problem
Find an elimination order that keeps the factors small
for the net below, or show that there is no such
order.
A
B
C
D
J
E
F
K
G
H
M
L
N
I
O
P
Lecture 16 • 73
Here’s a pretty complicated graph. But notice that no node has more than 2 parents,
so none of the CPTs are huge. The question is, is this graph hard for variable
elimination? More concretely, can you find an elimination order that results
only in fairly small factors? Is there an elimination order that generates a huge
factor?
73 The Last Recitation Problem (in this
lecture)
Bayesian networks (or related models) are often
used in computer vision, but they almost always
require sampling. What happens when you try to
do variable elimination on a model like the grid
below?
A
B
C
D
E
F
G
H
I
J
K
L
M
N
O
P
Q
R
S
T
Lecture 16 • 74
74
Lecture 16: Inference in Bayesian Networks
of 74
Recommended Documents
⏰ Deadline pressure?
Get your assignment done in just 3 hours. Quick, easy, and available 24/7.
Report
Tell us what’s wrong with it:
Thanks, got it!
We will moderate it soon!
Free up your schedule!
Our EduBirdie Experts Are Here for You 24/7! Just fill out a form and let us know how we can assist you.
Take 5 seconds to unlock
Enter your email below and get instant access to your document