fight my ignorance here, but it seems to me that a regular set of rules is isomorphic to a bayesian network with all rules and facts assigned a probability of 1.
If so, why not incorporate bayesian network inference into the standard commoditized rules engines and, possibly, still keep the regular inference as an optimization/special case module for faster handling of rule sets without probabilities?
Looking at this from another angle, given the engines and languages that are actually available now, how should the architect go about deciding which type of an engine should be used in a given project? What are advantages and disadvantages of bayesian inference vs regular inference engines?
This sounds like you’re adding in a layer of complexity for no benefit at all. That’s the dumbest thing a developer can do.
Rules-based decision engines are much easier and take less time (and hence cost less to develop). A bunch of “If X Then X” rules is a lot simpler than “Add X to the sum of all previous knowledge about Y, then recalculate the probability of Z.”
do you mean that the inference engine itself is hard to make? Or do you mean that I am ignorant in believing that making up a rule base with probabilities assigned (possibly not all that realistically in some cases) is about as easy as making a rule base without the probabilities?
AFAIU as an architect of a new system, I would treat both a logical inference engine and a bayesian network engine as a black box that magically generates inferences from whatever sets of rules and facts I put in. Is that an inaccurate assumption? Or is the bayesian one just harder to work with nowadays?
Again, what’s the benefit?
Magical black boxes are great when they work, but hard to debug when they don’t.
Yes. I mean exactly that. Bayesian engines are not simply “rule bases without probabilities.” I suggest the Wikipedia article on Bayesian probability as a starting point.