Configuration.check_valid_configuration() missleading documentation

[LukasFehring]

The function Configuration.check_valid_configuration() is documented to check whether a config is valid. However, it is not well specified when a configuration is valid.
Internally, the function calls:

from ConfigSpace.util import check_configuration

check_configuration(
  self.config_space,
  self._vector,
  allow_inactive_with_values=self.allow_inactive_with_values,
)

However, the function check_configuration does not appear to check whether the specified config is in the bounds of the ConfigurationSpace. It appears to mainly check whether the Conditions are met

Is this intentional?

[thijssnelleman]

This this question still applicable?

[LukasFehring]

what do you mean by apoplicable? I think it still exists

[thijssnelleman]

Just checking whether old issues are still relevant, will look into this

[thijssnelleman]

See new PR; this seems to be working fine in ConfigSpace as we are not allowed to create configurations that do not follow the ranges of the space. However, still checking whether this is also verified in the check_configuration function. I do wonder in what cases this could happen? I would like to adapt the new test accordingly to cover this.

[LukasFehring]

This is related to #412. If parameter bounds change, SMAC will still consider previously encountered data points in its hill-climbing algorithm. I tried to remove them based on whether they are legal or not. I then found that this is not working, even if the utilized configuration space object is changed.

[thijssnelleman]

That makes total sense. Once #412 has been merged, I will include in the test an adaptation of the parameter, and subsequent verification that (although the configuration exists but is no longer valid) will be rejected by the check function.

[mfeurer]

The assumption underlying the configspace is that a vector is mapped into the unit hypercube by the bounds given by the configuration space. Therefore, as long as the vector contains entries between 0 and 1, it contains valid entries. There is currently no mechanism in place that allows updating configurations (yet). Maybe these assumptions should be documented somewhere, for example, on a new documentation page outlining the design of the config space package.