Converting to Ordinary C# 7

In the previous lesson, we proposed a stripped-down DSL for probabilistic workflows. In this lesson, let’s see how we could “lower” it to ordinary C# 7 code. We will assume that we have all of the types and extension methods that we’ve developed so far.

Naive Approach

The first thing I’m going to do is describe a possible but very problematic way to do it.

We know how C# lowers methods that have yield return statements: it generates a class that implements IEnumerable, rewrites the method body as the MoveNext method of that class, and makes the method return an instance of the class. We could do the same thing here.

Recall that last time we gave three examples, the longest of which was:

probabilistic static IDiscreteDistribution<string> Workflow(int z)
{
  int ii = sample TwoDSix();
  if (ii == 2)
    return "two";
  condition ii != z;
  bool b = sample Flip();
  return b ? "heads" : ii.ToString();
}

We could lower this to:

static IDiscreteDistribution<string> Workflow(int z) =>
  new WorkflowDistribution(z);

sealed class WorkflowDistribution :
  IDiscreteDistribution<string>
{
  int z;
  public WorkflowDistribution(int z)
  {
    this.z = z;
  }
  public string Sample()
  {
    start:
    int ii = TwoDSix().Sample();
    if (ii == 2) 
      return "two";
    if (!(ii != z)) 
      goto start;
    bool b = Flip().Sample();
    return b ? "heads" : ii.ToString();
  }
  public IEnumerable<string> Support() => ???
  public int Weight(string t) => ???
}

We’d need to make sure that we did the right thing if z got modified during the workflow and the condition was not met of course, but that’s a small detail.

We could similarly lower the other two.

The good news is: we get the right distribution out of this thing:

Console.WriteLine(Workflow(3).Histogram());

    4|***
    5|****
    6|*****
    7|*******
    8|******
    9|****
   10|***
   11|**
   12|*
  two|**
heads|****************************************

The bad news is:

• We haven’t computed the support or the weights, as required.
• We are once again in a situation where a condition causes a potentially long-running loop to execute; we could do hundreds or thousands of iterations of going back to “start” for every successful sample if the condition was rarely met.

Basically, we’ve lost the great property that we had before: that we ...