Blake describes an enhancement of Quinlan’s system INFERNO for probabilistic inference. This enhancement is based on Hinton’s work in linear optimization. While “INFERNO allows only a very limited degree of propagation, as an essential precaution against non-termination of the inference process,” Hinton’s technique “allows effectively unlimited propagation; but termination is nonetheless assured by using a convergent algorithm. . . .”
The paper first describes Quinlan’s INFERNO, then Hinton’s constraint satisfaction system, then Blake’s marriage of the two. An abstract example is given, a summary is provided, and appendicies flesh out the details of the optimization algorithm and provide a real-world example.
The background required to properly appreciate this paper includes some knowledge of linear programming, artificial intelligence (AI), discrete mathematics, logic, and probability. The paper is well written, although a bit dense, and should interest researchers in AI, operations research, and applied logic alike.