compare to more algorithms

[KiaraGrouwstra]

I can either:

• compare NSPS/mine on other papers' benchmark functions (generating my own training data):
  • look for some new PBE paper with benchmark functions incorporating strong polymorphism without funky extras like recursion. this basically seems doubly unrealistic. 🙁
  • settle for the benchmark functions of an existing paper, which may push me into adding recursion (recursion #12), which I see as unnecessary added complexity for the purpose of my experiment and therefore if possible want to avoid Tamandu seems a viable candidate for! 🙂
• run other algorithms on a dataset of my choosing (designed to demonstrate the circumstances where I expect to add value). honestly though, comparing to type-only synthesizers may be unfair.
  • existing implementations (how hard will this be to integrate given for my own parts I've used Haskell?):
    • traditional:
      • deepproblog (github)
      • Typed MIL (code)
      • lambda^2 (code)
      • Synquid (code)
      • Tamandu (code)
      • Suggesting Valid Hole Fits for Typed-Holes in Haskell (code)
      • djinn (code) (+ related)
      • Speeding up Type-Driven Program Synthesis with Polymorphic Succinct Types ( code)
      • Idris (code)
      • rosette (github)
      • Live functional programming with typed holes (code)
    • neural (needs training):
      • NPI (github, pwc)
      • Alpha NPI (github)
      • deepcoder (github, pwc)
      • Leveraging grammar and reinforcement learning for neural program synthesis (code)
      • FrAngel (code)
  • re-implement algorithms without implementations:
    • traditional:
      • flashmeta or other Gulwani papers
      • Myth or other Osera papers like scout / scythe
      • 関係的仕様からの関数型プログラム合成
      • Automated Synthesis of Functional Programs with Auxiliary Functions
    • neural (needs training):
      • Neural-guided deductive search

TODO: categorize existing papers based on whether they use types, examples, or both.