Configuration Space does not allow Updates

[LukasFehring]

I believe that Configuration Space does not allow for online updates during its utilization. That would be very helpful, for example, in the case of Search Space Shrinking, or online control over optimization algorithms.

Concretely, it would be useful to allow:

1. Dynamically adapting the bounds of a hyperparameter
2. Dynamically adapting the distributions of hyperparameters

There might be other dynamic adaptations.

[LukasFehring]

Example code, that would be useful, but does not work currently

from ConfigSpace import ConfigurationSpace, UniformFloatHyperparameter

# Create a configuration space
cs = ConfigurationSpace()

# Add a single hyperparameter, e.g., learning_rate in [1e-4, 1e-1] (log-scaled)
lr = UniformFloatHyperparameter(
    name="example",
    lower=0,
    upper=1,
)

cs.add(lr)

# Example: sample one configuration
configs = cs.sample_configuration(5)
print("Before Adaptation:", configs)

hyperparameter = cs["example"]
hyperparameter.upper = 0.3
configs = cs.sample_configuration(5)

print("After Adaptation:", configs)

Output

Before Adaptation: [Configuration(values={
  'example': 0.6369527341018,
}), Configuration(values={
  'example': 0.217673016241,
}), Configuration(values={
  'example': 0.1592467284161,
}), Configuration(values={
  'example': 0.4795892828225,
}), Configuration(values={
  'example': 0.0662292258439,
})]
After Adaptation: [Configuration(values={
  'example': 0.7536177555621,
}), Configuration(values={
  'example': 0.3542359773483,
}), Configuration(values={
  'example': 0.2964701814732,
}), Configuration(values={
  'example': 0.9975855718142,
}), Configuration(values={
  'example': 0.6256529481526,
})]

[thijssnelleman]

This makes sense to me as a feature, but there could be several implications for this change. Will look into this next month.

[LukasFehring]

I briefly discussed this with @eddiebergman at some point. I remember him saying that it is mainly a cachin issue.

[LukasFehring]

I would also really like to be able to change the type of a Hyperparameter. For example, from Float, to Categorical :) This would be an extreme space of search space shrinking :)

[thijssnelleman]

We are currently working on the removal of HPs, I will look what is necessary/overlapping with this

[thijssnelleman]

First PR version; take a look

#414

[LukasFehring]

Hey, thank you so much for the update. This already looks really cool. Does this also tackle the neighborhoods? https://automl.github.io/ConfigSpace/latest/reference/hyperparameters/#neighborhoods.

This is a small code block copied from https://github.com/automl/SMAC3/blob/main/smac/acquisition/maximizer/local_search.py. At first glance, I am unsure if this works with the adaptations in PR.

 for i, inc in enumerate(candidates):
            neighborhood_iterators.append(
                # get_one_exchange_neighbourhood implementational details:
                # https://github.com/automl/ConfigSpace/blob/05ab3da2a06c084ba920e8e4e3f62f2e87e81442/ConfigSpace/util.pyx#L95
                # Return all configurations in a one-exchange neighborhood.
                #
                #     The method is implemented as defined by:
                #     Frank Hutter, Holger H. Hoos and Kevin Leyton-Brown
                #     Sequential Model-Based Optimization for General Algorithm Configuration
                #     In Proceedings of the conference on Learning and Intelligent
                #     Optimization(LION 5)
                get_one_exchange_neighbourhood(inc, seed=self._rng.randint(low=0, high=100000))
            )
            local_search_steps[i] += 1