Consistency of Findings

There are three ways that findings can be inconsistent:

1. Several findings for different nodes can be inconsistent with each other.  This is the most common form of inconsistency.  When Netica alerts you that findings are inconsistent, then it means that according to this Bayes net, it is impossible to observe the set of findings you have entered.  If you know that those findings are possible, then the fault lies in the net’s conditional probabilities (see below).

2. Several findings for the same node can be inconsistent with each other.  This only applies to findings entered in accumulation mode, because otherwise each new finding will retract the previous one for that node.  Usually only negative or likelihood findings are entered in accumulation mode, and the only way they can be inconsistent is to have at least one negative finding (or zero likelihood) for every state of the node.

3. A single finding can be inconsistent with the net itself.  This is rare.  Basically the net has a zero prior probability for that finding, which means it is a finding which should never occur.  Usually this indicates the net was not designed to handle cases of this type.

When Detected:  Netica will always detect an inconsistency of type 2, and usually of type 3, as soon as you try to enter it.  During belief updating all inconsistencies will be detected, and Netica will report them (either to the Messages window or with a dialog box).  So if the beliefs are kept current by belief updating after each finding is entered (which is one reason to use auto-updating), then you will be alerted as soon as you try to enter an inconsistent finding, otherwise you won’t be alerted until the next belief updating is in progress.

Recovery:  If Netica reports an inconsistent finding, simply remove it and continue.  Nothing within Netica will be left in a problematic state.

Before Entering:  You can always tell if a finding will be inconsistent before you enter it by looking at the belief for that state (do belief updating first if necessary).  If the belief is zero, the finding will be inconsistent, otherwise it will be consistent.

Incorrect Net:  If you have problems with consistency, re-examine whether conditional probabilities you have entered as 0% should really indicate absolutely impossible.  If there is some very small probability that they aren't impossible, then enter that very small probability instead.  Equivalently, perhaps some nodes that are deterministic should really be probabilistic.  Also consider the possibility that the net is correct, but that the findings should be high likelihood findings, rather than positive findings.

Inconsistent Net:  A net can never be inconsistent with itself.  No matter what its structure, or its conditional probability tables, it always represents some joint probability distribution.  Only findings can be inconsistent.  However, a net may be incorrectly designed, as described above.