How BBNs deal with evidence
In this section we look at the way that evidence is transmitted in BBNs. We
consider two types of evidence:
- Hard evidence (instantiation) for a node X is evidence that the state of X is definitely a particular value.
For example, suppose X represents the result of a particular match for a
football team {win, lose, draw}. Then an example of hard evidence would be knowledge
that the match is definitely won. In this case we also say X is instantiated as the value 'win'.
- Soft evidence for a node X is any evidence that enables us to update the prior probability
values for the states of X. For example, if we know that the team is winning
the match 3-0 at half-time, then the probability of win would be quite high, while the probabilities of both lose and draw would be quite low (compared with ignorance prior values).
We distinguish three types of connection in a BBN:
In a serial connection we will see that any evidence entered at the beginning of the connection can
be transmitted along the directed path providing that no intermediate node on
the path is instantiated (which thereby blocks further transmission).
In a diverging connection we will see that evidence can be transmitted between two child nodes of the
same parent providing that the parent is not instantiated.
In a converging connection we will see that evidence can only be transmitted between two parents when
the child (converging) node has received some evidence (which can be soft or
hard).
The rules for transmitting evidence for serial, diverging and converging
connections are sufficient for us to describe a completely general procedure for
determining whether any two nodes of a BBN are dependent or not. This is the
formal notion of d-separation. This notion is crucial for understanding how the algorithms for probability
propagation in BBNs actually work.