Refactor the CBBO and DTLZ benchmarks

[Deathn0t]

The DTLZ benchmarks should be refactored to follow the new setup.

• source: https://github.com/deephyper/benchmark/tree/0.0.1/lib/DTLZ
• example: https://github.com/deephyper/benchmark/tree/refactoring/src/deephyper_benchmark/lib/c_bbo
• branch: refactoring

[wigging]

Is uv or conda used for development of this benchmark package?

[Deathn0t]

I used uv sync from the root of this repository.

[Deathn0t]

A benchmark will now be defined as a subclass of HPOBenchmark (see definition)

An example usage will be:

from deephyper_benchmark.benchmarks.cbbo import AckleyBenchmark as bench

search = CBO(bench.problem, bench.run_function)
results = search.search(max_evals=100)

• cbbo
  • ackley @Deathn0t
  • branin @Deathn0t
  • easom @Deathn0t
  • griewank @Deathn0t
  • hartmann6D @Deathn0t
  • levy @wigging
  • michal @wigging
  • rosen @wigging
  • schwefel @wigging
  • shekel @wigging
• dtlz @Deathn0t

The logic to add noise to the objective or to add a sleep time to the objective can be factoried as deephyper_benchmark.utils decorators noisy and sleep.

Each Benchmark should also have a property scorer that returns a subclass of HPOScorer (see definition) with methods that transform the objective in regret.

[wigging]

How is this so-called "scorer" used? Can you point me to an example in the old CBBO and DTLZ code https://github.com/deephyper/benchmark/tree/0.0.1? In the old code I see a metrics.py module that has regret functions but I don't see any examples of this being used. Also, what does "regret" mean in this context, is it supposed to be regression?

[Deathn0t]

I added a run_function(..., bb_func=...) at deephyper_benchmark.benchmarks.cbbo.utils to factorize the logic:

e7293e2

[wigging]

I added the remaining C-BBO benchmarks. Can we start a pull request for the cbbo branch to review the changes before merging with the main branch?

[Deathn0t]

I merged the CBBO branch in #34 after editing.