Bayesian Network Models

Much of the power and innovation offered by MODIST stems from its animation of process models using Bayes' theorem for probabilities. This allows us to coherently revise the probabilities of events as evidence becomes available.

Consider a hypothesis or event A that is conditional on evidence provided by another event B. A and B have individual prior probabilities p(A) and p(B). The likelihood of observing B if A is true is given by p (B|A). Bayes' theorem gives the posterior probability for A as:

Of great practical significance is that we can reason in two directions: forwards from cause to effect, and backwards from a desired or observed effect to its possible causes. So we can reason quantitatively in either deductive or adductive mode. We can choose whichever best addresses the situation given the information we have available. This is another strength of causal models compared with statistical models.

MODIST encapsulates the variability in the software process within the nodes in the process models. Typically, we shall have evidence for some variables, and we want to predict the others. A feature of Bayesian Networks is that they consistently propagate the impact of evidence among the probabilities of all the uncertain outcomes. So, entering new evidence at any node updates the probabilities associated with all the other nodes in the network.

To distinguish the application of BN models to process control, MODIST uses the term Bayesian Process Control (BPC). MODIST has made significant advances in implementing BN models to support BPC.