epsilon greedy candidate selector

[vacmar01]

I implemented a epsilon greedy candidate selector strategy, added it to the API and added some tests (for all candidate selectors).

It is not fully documented yet, but I wanted to make sure that I'm on the right path with the API. I changed it a bit so that optimize also accepts a CandidateSelector subclass directly, so that we can specify further hyperparameters (like the epsilon in this case).

What do you think, @LakshyAAAgrawal ? :)

Plan is to add more candidate selection strategies (Softmax, Thompson Sampling, UCB1 etc.) to the repo and compare them to each other.

[semanticdiff-com]

Review changes with  

Changed Files

File	Status
src/gepa/api.py	42% smaller
src/gepa/strategies/candidate_selector.py	0% smaller
tests/test_candidate_selector.py	0% smaller

[fgn]

Consider restricting ε-greedy exploration to a more refined pool of candidates. Currently, exploration may select any candidate, including those that are clearly suboptimal.

You might parameterize EpsilonGreedyCandidateSelector with a pool argument accepting values like "pareto", "top_k", or "all", along with an optional k parameter for the "top_k" setting.

One suggestion could be to start with the Pareto set to ensure a good balance between exploration and exploitation, and then experiment with the top_k setting to fine-tune performance.

[fgn]

Don't call it _cls here when it's not a class.

[fgn]

This will instantiate all three selectors.

Here is a lazy pattern that will instantiate at most one.

factories = {
    "pareto": lambda: ParetoCandidateSelector(rng=rng),
    "best": lambda: CurrentBestCandidateSelector(),
    "epsilon_greedy": lambda: EpsilonGreedyCandidateSelector(epsilon=0.1, rng=rng),
}

try:
    candidate_selector = factories[candidate_selection_strategy]()
except KeyError:
    raise ValueError(
        f"Unknown candidate_selector strategy: {candidate_selection_strategy}. "
        "Supported strategies: 'pareto', 'best', 'epsilon_greedy'"
    )

[vacmar01]

Great review, thanks :) Will incorporate the changes.