New StringVariable

[eduardo-ocampo]

Adds a new StringVariable type for carrying string values through a problem setup. Closes #56

Because strings are not something an optimizer can act on, StringVariable types are supported as input or an output only and always treated as fixed with an OptProblem.

Changes:

• New StringVariable inherited from FloatVariable and modeled after CatgoricalVariable
• Added StringVariable to Variable union. As well as being importable via standard_evaluator.StringVariable
• Created OptProblem validation for StringVariable. Errors if StringVariable is used as an objective or constraint.
• Modified build_maps utility to treat variables ( StringVariable) with no numeric bounds as fixed.
• For StringVariable, calculate_default method does not overwrite an existing default since it has no bounds to derive from. StringVariable.calculate_default is maintained to be consistent with the other variable types. Only fills with empty string if no default was set.
• abstract_evaluator.__call__'s fixed variable filter now guards against bounds being None

Testing

• Adds tests in tests\test_problem.py to construct and validate StringVariable. Additional tested created to test StringVariable interaction with OptProblem
• New test tests\test_string_variable_evaluation.py asserts wrapped evaluator matches a direct call and that the types are preserved.

[jmgablon]

Hi Eduardo,
this looks really good. I like that we use the existing capabilty to mark something as fixed to deal with string variables.

I have some comments, especially that a string variable can be an output.

We also need to make sure that string inputs or outputs are not overwritten when running an evaluation on a DataFrame. We should have a test that checks what happens with this:

import pandas as pd

df = pd.DataFrame({'input': ['one', 'two'], 'val': [3.4, 5.6]})

def my_example(df: pd.DataFrame) -> None:
    df['out'] = df['input'] + df['val'].astype(str)  # Creates a string output
    df['val_out'] = df['val']*4.

When we wrap this as an evaluator we want to get the same output as if we just call the function directly.

[jmgablon]

So we want to allow string variables as responses. Sometimes the output of an analysis can for example be a unique file name that points to a file that was generated by the analysis.

[eduardo-ocampo]

Changes were made in recent commits to allow StringVariables as responses.

[jmgablon]

I don't think we should allow an empty name. Let's examine this in a different MR.