Is this comparable to causal networks in some sense, other than being represented as a directed graph? Digraphs are ubiquitous in applications of math, they’re used for all sorts of things.
I’m not sure if this is what jimrandomh was getting at, but there does seem to be a connection to me deeper than merely both being DAGs: optimizations are often limited by how much inference can be made about the causal relations of variables and functions.
One of Haskell’s purposes is to see what happens when you make dependencies explicit, removing most global state, and what optimizations are enabled. For example, consider constant propagation: a constant is a node unconnected to any other node, and therefore cannot change, and being unchanging can be inlined everywhere and assumed to be whatever its value is. If the constant were not constant but had a causal connection to, say, a node which is an IO String that the user types in, then all these inlinings are forbidden and must remain indirections to the user input.
Or consider profile-guided optimization: we relocate branches based on which one is usually taken (probability!), but what about branches not specifically in our profiling sample? Seems to me like causal inference on the program DAG might let you infer which branch is most likely.
Is this comparable to causal networks in some sense, other than being represented as a directed graph? Digraphs are ubiquitous in applications of math, they’re used for all sorts of things.
I’m not sure if this is what jimrandomh was getting at, but there does seem to be a connection to me deeper than merely both being DAGs: optimizations are often limited by how much inference can be made about the causal relations of variables and functions.
One of Haskell’s purposes is to see what happens when you make dependencies explicit, removing most global state, and what optimizations are enabled. For example, consider constant propagation: a constant is a node unconnected to any other node, and therefore cannot change, and being unchanging can be inlined everywhere and assumed to be whatever its value is. If the constant were not constant but had a causal connection to, say, a node which is an
IO Stringthat the user types in, then all these inlinings are forbidden and must remain indirections to the user input.Or consider profile-guided optimization: we relocate branches based on which one is usually taken (probability!), but what about branches not specifically in our profiling sample? Seems to me like causal inference on the program DAG might let you infer which branch is most likely.