search

racket / redex

Other
93 stars 36 forks source link

documentation: what are generation failures #201

Open wilbowma opened 5 years ago

wilbowma commented 5 years ago

I have a program using redex-check that looks like


> (redex-check
    BoxyTypingL
    #:satisfying (type-infer · · e A)
    (redex-match? BoxyEvalL v (term (eval e)))
    #:attempts 1000)
redex-check: no counterexamples in 78 attempts (with 922 generation failures)

I cannot for the life of me figure out why there would be so many failures, and looking through the documentation, I don't see any references that help me make sense of it.

As I understand #:satsifying, there ought not be any failures! It should have generated a bunch of well-typed terms...

rfindler commented 5 years ago

There are a lot of reasons why there might be generation failures. I mean, basically, it is a randomized solver that is going off and searching for solutions and sometimes it backs itself into a corner and can't get out. So it gives up and restarts the process. One thing that can happen (that is hard to diagnose) is simply that randomly selecting the next "layer" of the derivation has a relatively low probability of success. For example, if there are 10 choices and making any one choice means that you need to select a specific next choices, well, that's effectively 90% chance of failure (that you'll repeat many times as you build up the tree). That's the main issue, but there are others too, like the generator might just not be smart about certain combinations of redex features somehow in ways that could be fixed but we just haven't identified yet and, of course, there could be a bug! (I'm pretty sure that historically there were just regular ole bugs that manifested as certain models always or almost always failing generation.)

rfindler commented 5 years ago

Oh, and if someone actually wants to get dirty working with the solver, I believe that @bfetscher built something to visualize the steps that the solver takes so you can try to get a feel for what's going on with a specific model in an attempt to improve things. I've not really used it myself, however.

rfindler commented 5 years ago

Oh, and @bfetscher gave a nice talk about how this all works but I cannot find it online. I stole his slides and gave a version of it here: https://www.youtube.com/watch?v=SEb7FjZwRUk in case you're curious about how the solver is working (although this may already be old news, in which case I apologize).

wilbowma commented 5 years ago

Okay this looks like Redex is not doing what I sort of expected, or is counting attempts differently than I expect. I was sort of imagining that Redex would do more or less the same thing as a miniKanren program, so if I request 1000 solutions, it would search backwards until it found 1000 valid solutions. It sounds like Redex instead searches until it has tried 1000 unifications (or something), some of which fail.