Active Trails © Prof. Ross D.

Shachter

Active Trails
A belief network (or relevance diagram) is a graphical model used to represent the
structure of the relationships among uncertain variables. The diagram is built based on
the decision maker’s background state of information, S, which might or might not be
explicitly shown in the diagram. A node is the basic element in a belief network. Each
node corresponds to a specific uncertain variable. We will refer to the node and the
corresponding variable interchangably. For example, in Figure 1, there are uncertain
nodes X, C, Y, A, E, B, and D, in addition to the background state of information S.

X C Y

A E

S B D

Figure 1. Sample Belief Network

An arc is an arrow pointing from one node to another node in the diagram. We refer to
the sending node as a parent, the receiving node as a child, and a node’s parents and
children as its neighbors. We say that the arc is conditional because it indicates that a
probability distribution for the child will be assessed conditioned on its parents and S.
For example, in Figure 1, X and C are the parents of A and A is a child of X and C.
We do not permit any directed cycles in the diagram, a loop where we can return to a
given node by following a path in the direction of the arcs. As a result we are
guaranteed that there is an assessment order of all of the nodes in the diagram in which
all of the parents of each node X precede X. For example, in Figure 1, one possible
assessment order is S, X, C, Y, A, E, B, D. Another is S, Y, C, E, D, X, A, B.
Judgments of independence are established as we build a diagram in assessment order
and decide which of the nodes that precede it should be parents. For each node X, we
are asserting that if we are able to observe its parents then we can learn nothing more
about X by observing the other nodes Z that precede X in the order. We can state this as
“X is irrelevant to Z given the parents of X and S.” For example, in Figure 1, we know that
B must be independent of X, C, and Y given A, E, and S.
Once the diagram has been built we can recognize additional statements of
independence that must follow from the judgments that we made in assessment oder.
We can recognize these inferred independence statements by a graphical technique
applied to the belief network. We are going to look for trails, sequences of neighboring
nodes that don’t necessarily follow the directions of the arcs. For example, in Figure 1,
trails include X-A, X-A-B-E-Y, and X-A-C-E-D-E-Y. The reverse sequences, A-X, Y-E-B-
A-X, and Y-E-D-E-C-A-X, are also trails.
For each node on the trail, we look at the direction of the arcs between it and the nodes
that precede and follow it on the trail. We say that it has two converging arcs if it is an
intermediate node on the trail (neither the first nor the last), and both nodes on either
side in the sequence are its parents. For example, in Figure 1, for the trail X-A-B-E-Y, B
is the only node with two converging arcs. For the trail X-A-C-E-Y, both A and E have
Active Trails © Prof. Ross D. Shachter

two converging arcs. Finally, for the trail X-A-C-E-D-E-Y, A and E are the nodes with
two converging arcs, where D’s parent, E, both precedes and follows it on the trail.
We are interested in determining whether the belief network asserts that two nodes,
such as X and Y, must be independent given a (possibly empty) set of nodes Z and S.
That is, having observed Z, whether we cannot learn anything more about X by
observing Y. We say that there is an active trail between X and Y given Z if there is
some trail between X and Y such that every node on the trail with two converging arcs
is observed and none of the others are observed. For example, Figure 2 shows multiple
examples of active trails between X and Y, where observed nodes are shaded. Both
shaded nodes have two converging arcs on the trail between X and Y but none of the
other nodes do. The active trails are X-Y for a, X-A-Y for bcde, and X-A-B-A-Y for f.

X b) A X Y
a)
c)
X Y A
Y

X e) Y X Y

d) f)
A A A

Y X B

Figure 2. Active trails between X and Y, where observed nodes are shaded
Figure 3 shows multiple examples where there are no active trails between X and Y.
The inactive trails are X-Y for abc and X-A-Y for defg. The observed nodes are shaded
and none of them have two converging arcs, so they prevent trails from being active for
abcdef. The only node with two converging arcs is A in g, but it is not observed and thus
prevents trail X-A-Y from being active for g.

X X X d) A

a) b) c)
X Y
Y Y Y

X f) Y X Y

e) g)
A A A

Y X

Figure 3. None of the trails between X and Y are active

Active Trails © Prof. Ross D. Shachter

Now that we can recognize when there are active trails between X and Y given Z, we
can take our final step. If there is such an active trail then X and Y might be dependent
given Z and S. If there is no such trail, then X and Y must be independent given Z and S.
For example, in Figure 2, X and Y might be dependent given S for abde, X and Y might
be dependent given A and S for c, and X and Y might be dependent given B and S for f.
In Figure 3, X and Y must be independent given X and S for a, given Y and S for b, given
X, Y and S for c, given A and S for def, and given S for g.
For any node in a belief network, which nodes are independent of it given its parents
and S? If the diagram in Figure 1 was constructed using assessment order S, X, C, Y, A,
E, B, D, we know right away that A is independent of Y given X, C, and S. However,
the only active trail from A given X and C is A-B. Therefore, A is also independent of E
and D given X, C, and S. It turns out that any possible assessment order, such as S, X, C,
Y, E, D, A, B, would give us the same belief network. In general, using active trails, we
can see that every node is independent of any node that is not its descendant, given its
parents and S. In this case, A is independent of Y, E, and D, given X, C, and S.
Likewise, X is independent of C, Y, E, and D given S, and B is independent of X, C, Y,
and D given A, E, and S.

Practice Problems based on Figure 1

1. In Figure 1, are X and A independent
a) given S?
b) given B, C, and S?
c) given A and S?
2. In Figure 1, are X and B independent
a) given S?
b) given A and S?
c) given A, C, and S?
d) given A, E, and S?
3. In Figure 1, are X and Y independent
a) given S?
b) given A and S?
c) given B and S?
d) given A, E, and S?
e) given A, B, and S?

Practice Problem Solutions are on the next page.

Active Trails © Prof. Ross D. Shachter

Practice Problem Solutions

1. a) X-A is an active trail so X and A might be dependent given S
b) X-A is an active trail so X and A might be dependent given B, C, and S
c) There is no active trail between X and A given A, so X and A must be
independent given A and S
2. a) X-A-B is an active trail so X and B might be dependent given S
b) X-A-C-E-B is an active trail so X and B might be dependent given A and S
c) There is no active trail between X and B given A and C, so X and B must
be independent given A, C, and S
d) There is no active trail between X and B given A and E, so X and B must
be independent given A, E, and S
3. a) There is no active trail between X and Y, so X and Y must be independent
given S
b) There is no active trail between X and Y given A, so X and Y must be
independent given A and S
c) X-A-B-E-Y is an active trail so X and Y might be dependent given B and S
d) X-A-C-E-Y is an active trail so X and Y might be dependent given A, E,
and S
e) X-A-C-E-B-E-Y is an active trail so X and Y might be dependent given A,
B, and S