A system for exact inference on relational Bayesian networks, compiled into arithmetic circuits and answered in time linear in the circuit's size, on instances with thousands of variables.
Relational Bayesian networks describe a whole family of models at once, in the way the Primula tool defines them. The paper describes a system that compiles propositional instances of these networks into arithmetic circuits, then answers queries by evaluating and differentiating the circuit in time linear in its size. It reports successful compilation and efficient inference on instances with thousands of variables, whose jointrees have clusters of hundreds of variables.
Compiling a model once and then answering by evaluating a circuit is the pattern the rest of this record rests on. Here it is shown working at a scale where the standard jointree method has no room left.
A methods paper on relational models. No product claims attach to it.