Conflict is a measure that detects evidence that is conflicting or rare. The greater the conflict value above zero, the more likely the evidence is in conflict, or rare.

Conflict = log((P(e1)P(e2)...P(ei)) / P(e)), where P(e1), P(e2) etc... are the likelihoods of each variable considered in isolation, and P(e) is the likelihood of the all the evidence together (see Log likelihood ).

If evidence on a variable X agrees with evidence on another variable Y, then often P(X=x)P(Y=y) < P(X=x, Y=y), and will therefore lead to a negative conflict.

Bayes Server supports conflict on networks with discrete and/or continuous variables, and also dynamic Bayesian networks. However while calculations involving discrete variables are fast, there is additional overhead for continuous variables.