search

KiaraGrouwstra / synthesis

code-base for my UvA MSc AI thesis on typed neural program synthesis
https://gitlab.com/KiaraGrouwstra/hasktorch/tree/synthesis/synthesis/
Other
5 stars 1 forks source link

recursion #12

Open KiaraGrouwstra opened 4 years ago

KiaraGrouwstra commented 4 years ago

like every benchmark set I considered incorporates function recursion: tamandu, lambda^2, myth. I fear if I wanna be able to use their benchmark tasks as-is, I'll need to augment my grammar to account for this, i.e. go from an expression of combinators to a (recursive) function definition (add Match, FunBind as per grammar), which also requires variables (#8) due to the function parameters.

I'm kind of unhappy about this though, as you may need types to work with recursion. while that may not so much be an issue in synthesis, where you may know the target type signature, this feels tougher in task function generation for my training set, where I'm currently just generating any seemingly sensible function without regard to types. maybe it'd just work -- I should try.

KiaraGrouwstra commented 4 years ago

figuring I can also compare to other papers using my own dataset (#13), this now seems like a nice-to-have rather than a critical feature.