feat(TRIM): add support for entities that are missing the target variable measurement

[VassilisVassiliadis]

Is your feature request related to a problem? Please describe. A clear and
concise description of what the problem is. Ex. I'm always frustrated when [...]

The current version of TRIM requires that your experiment can obtain a target variable measurement for every entity in the space and that these measurements are of the same type (numbers or strings, vectors are not supported).

Describe the solution you'd like. A clear and concise description of what
you want to happen.

It is difficult to design a one-size-fits-all solution. So we'll support 3 mutually exclusive modes of operation when using TRIM to apply the experiment on an entity which does not measure the TRIM target variable:

• (default) Raise an exception (current behaviour)
• Use a default value, with up to missingTargetMeasurements.budget times and raise an exception when there's too many measurements that are missing the target variable
• Skip an Entity, up to missingTargetMeasurements.budget times. Raise an exception if skipping too many entities OR if running out of entities

The default behavior is to raise an exception on first missing target variable measurement. The default of missingTargetMeasurements.budget is None which means that the budget is infinite. The missingTargetMeasurements.budget parameter of TRIM must either be None, or greater than 0.

To pick a mode set the parameter missingTargetMeasurements.mode which can be one of:

• RaiseError
• InjectDefaultValue - which uses the value of missingTargetMeasurements.defaultValue
• Skip

if missingTargetMeasurements is not set, then it defaults to missingTargetMeasurements.mode = RaiseError

We'll also refactor the NoPriorsSampleSelector to keep track of whether the operator managed to measure the target variable or not.

Describe alternatives you've considered. A clear and concise description of
any alternative solutions or features you've considered.

• Update the actuators to use a default value instead: This is a hack

Additional context. Add any other context or screenshots about the feature
request here.

[VassilisVassiliadis]

There are 2 samplers NoPriorsSampleSelector for the sampling of the N initial points and TrimSampleSelector for the iterative sampling of points while the model is being created.

NoPriorsSamplesSelector does not check whether the Operation that's calling the sampler (i.e. RandomWalk) successfully records the value of the target variable for the entity that the sampler is yielding. TrimSampleSelector does, but not directly. We need to update TRIM to insert the default value when using either of the samplers and that requires a bit of careful thinking first.

The TrimSampleSelector keeps track of the dataframe that represents the space before and after it yields an entity to the Operation that's invoking the sampler (i.e. RandomWalk). This sounds like a workaround.

The right way to go about this is to figure out the right way to track whether TRIM got a Measurement or an InvalidMeasurement when applying an experiment to an entity. Then, refactor TRIM to use this mechanism, and then enhance TRIM with a new feature for using an optional default value in the event that an experiment does not measure the target variable for a given entity.

[VassilisVassiliadis]

I'm putting together a quick PR that only patches the TrimSampleSelector sampler to use a default value for both Measurement and InvalidMeasurement scenarios where RandomWalk does not lead to observing a value for the target variable for the Entity that the sampler last emitted.