What We Have So Far

As per our meeting on 02/04/18, we believe the best way to work is through formalized notions of Domains and Solvers. These are abstract base classes for any domain or solver we might implement to promote compatibility between arbitrary solvers and domains.

Domains consist of:

• DSLs: {Class₁, ..., Class_N}

  • A DSL consists of a set of classes:

    DSL = {Add, Substract, Transpose, ...}

    Where Add, Subtract, and Transpose are all defined Python classes as described below.

• Class: {Signature, Func(LHS, RHS), Token}

  • A class within a DSL will consist of a signature (its name), a function, and a token that is able to be parsed:

    class Plus:
    # The class name is the signature
      def __init__(self):
        # The token
        self.token = '+'
      def __repr__(self):
        return self.token
      def f(LHS,RHS):
        # The function
        return LHS + RHS 

• Execute: Prog -> Spec

  • Execute(Prog, Spec) will take both a Program and a Spec and execute a Solver to solve the Domain

• Context: Spec* -> Ctx

  • A Context takes a Spec* and outputs a Ctx

• Loss: Spec -> Prog -> Real Number

  • Loss takes a Spec and Program and outputs a real number detailing how close the current Program and Spec are to the Spec*

Solvers consist of:

• Solver: DSL -> Spec -> Loss -> Timeout -> [(prog_N, loss_N, time_N)]
  • A solver will take a DSL, Spec, Loss, and a Timeout and output a list of size N of triples of Programs, Losses, and Times
  • Timeout is the maximum amount of time the solver has to solve the Spec

  • class Solver:
      def __init__(self, dsl, spec, loss, timeout):
        self.dsl = dsl
        self.spec = spec
        self.loss = loss
        self.timeout = timeout
        self.report_list = []
      def __repr__(self):
        # Each solver must have a representation (e.g. return "MCTS")
        pass
      def execute(self, program, spec):
        # Executing the solver
        pass
      def report(self):
        # Adding to report_list
        pass

Actual Domains:

• 2D CSG
• 3D CSG
• Vector Graphics
• Symbolic Regression
• Flashfill
• Theory Learning
• Tangrams
• Body Modeling

Actual Solvers:

• Random
• Enumeration
• DeepCoder
• Genetic Programming w/ Speciation
• Genetic Programming w/o Speciation
• EC
• Learn To Search