[Possible Minor Bug] Encoding of NaN's in run encoding

[eddiebergman]

I found this comment here where it says "NaNs should be encodere with -0.5" yet the actual imputation value is -0.2.

In def encode_config()

DeepCAVE/deepcave/runs/__init__.py

Lines 735 to 748 in daf6c2f

	for value, hp in zip(values, hps):
	# NaNs should be encoded as -0.5
	if np.isnan(value):
	value = NAN_VALUE
	# Categorical values should be between 0..1
	elif isinstance(hp, CategoricalHyperparameter):
	value = value / (len(hp.choices) - 1)
	# Constants should be encoded as 1.0 (from 0)
	elif isinstance(hp, Constant):
	value = CONSTANT_VALUE
	x += [value]
	return x

In constants.py

DeepCAVE/deepcave/constants.py

Lines 1 to 8 in daf6c2f

	NAN_VALUE = -0.2
	NAN_LABEL = "NaN"
	VALUE_RANGE = [NAN_VALUE, 1]
	CONSTANT_VALUE = 1.0
	BORDER_CONFIG_ID = -1 # Used for border configs
	RANDOM_CONFIG_ID = -2 # Used for random configs
	COMBINED_COST_NAME = "Combined Cost"
	COMBINED_BUDGET = -1

Not sure which is "correct". From the configuration footprint point of view (where I found it from), I imagine -0.5 is a bit better as everything else is scaled from (0, 1) when valid. However this encoding is a bit more general and used for other purposes it's probably good to find the original source to clarfiy. I don't know what that orginal source is unfortunately.