fix(rng): Ensure reproduciblity through construction and reset

[eddiebergman]

Heyo,

This address issue #83 which seemed a bit more deep-rooted than I thought, i.e. not just seeding the configspace.
The main change is to never share the rng object itself but rather pass around seeds generated from it, making things like reset() much easier to deal with. Also prevents having to save and store random state object internals.

Tested with the following script, which should probably be made into a full-fledged test for reproducibility.

from __future__ import annotations

from ConfigSpace import ConfigurationSpace

from dehb import DEHB

cs = ConfigurationSpace({"a": (1.0, 10.0)})


run1 = []
run2 = []
run3 = []
run_4 = []
run_list = run1


def f(config, fidelity):
    run_list.append((config["a"], fidelity))
    return {"fitness": config["a"] ** 2, "cost": 1.0, "info": {}}


run_list = run1
D = DEHB(cs, f=f, seed=1, min_fidelity=1, max_fidelity=100, n_workers=1, output_path="test1")
D.run(10)

run_list = run2
D2 = DEHB(cs, f=f, seed=1, min_fidelity=1, max_fidelity=100, n_workers=1, output_path="test2")
D2.run(10)

run_list = run3
D.reset()
D.run(10)

# Run it again to make sure its different than run3
run_list = run_4
D.run(10)


print(run1)
print(run2)
print(run3)
print(run_4)

[Bronzila]

Hey,
I just took some time to look over the changes and they look great! There is however a little extra work needed in order to properly allow restarting. I will try to incorporate these chaneges over the weekend :)

PS: I've redirected the PR to development, but will make sure that it will be part of the next release, thanks for your time :)

[eddiebergman]

Nice, sorry I didn't know where to develop off or include. Feel free to close this PR and just integrate the changes into whatever new version you make if it's easier!