What I don't understand is _what_ these applications are.
What does this surprise value tell me, and how can it be calibrated to anything quantitative? What's the equivalent of statistics that I can use to alter my original model's rankings to conform to what I observe?
I can understand ranking as coarse-graining of probability if there's some way to approximately map it back to the standard notion of probability.
On the other hand, using rank to direct collections of traces in a well-defined order is admittedly a neat trick; I can see using it to find something like the "most reasonable" or "easiest to explain" solution to a given logic problem. (I have actually seen a hand-coded version for Sudoku that attempts to minimize the backtracking depth to make it as easy to explain as possible.) But when would I want to do that to a program where the ranks represent instead how common a given event is? Verifying the no-exception case first? Is that useful?
I can understand ranking as coarse-graining of probability if there's some way to approximately map it back to the standard notion of probability.
On the other hand, using rank to direct collections of traces in a well-defined order is admittedly a neat trick; I can see using it to find something like the "most reasonable" or "easiest to explain" solution to a given logic problem. (I have actually seen a hand-coded version for Sudoku that attempts to minimize the backtracking depth to make it as easy to explain as possible.) But when would I want to do that to a program where the ranks represent instead how common a given event is? Verifying the no-exception case first? Is that useful?