feat (algo): Add MAP-Elites Low-Spread

[btjanaka]

Related issues: [refer to issues] Resolves #151

[Introduce the overall change made in the PR and list the associated modifications below]

Add MAP-Elites Low-Spread: https://dl.acm.org/doi/abs/10.1145/3583131.3590433

This PR introduces:

• MELSRepertoire in mels_repertoire.py
  • Tests in mels_repertoire_test.py
• MELS in mels.py
  • Tests in mels_test.py
• Spread type in types.py
• multi_sample_scoring_function in sampling.py

Questions

1. Should we store all n_evals fitnesses and descriptors in the repertoire, or should we just store the aggregated fitness and descriptors? We can also store both the aggregate and the separate values. Storing separate values would require passing in n_evals to MELSRepertoire.init_default so that we can define the shapes of fitnesses and descriptors.
2. Should we check if one individual dominates the others if inserting multiple individuals into the same cell?
3. What link should I use for citations? Currently I am using the ACM Digital Library link. I can also use the arXiv link: https://arxiv.org/abs/2303.16207 if preferred.

Checks

• a clear description of the PR has been added
• sufficient tests have been written
• relevant section added to the documentation
• example notebook added to the repo
• clean docstrings and comments have been written
• if any issue/observation has been discovered, a new issue has been opened

Future improvements

[List here potential observations made and/or improvements that could be made in the future. If relevant, open issues for those.]

[Lookatator]

Hi @btjanaka ! Thank you very much for this (super well-documented) PR.

I added @Egiob and @manon-but-yes as reviewers for this PR; they are probably the best people to judge it.

I had an overview of the code and it looks great! I have mainly one high-level comment regarding the implementation of the reevaluation scoring function (and where it should be implemented). I will add it as a comment.

[btjanaka]

Hi @Lookatator, thank you for taking a look at this PR! I addressed your comment on the pairwise distance calculation, but I do not see the comment on scoring function evaluation. Did you make the comment in the tests or the notebook? (the scoring function is in both those places).

[Egiob]

Hi guys! I will do a pass by the end of the week 😄

[Lookatator]

Hi @Lookatator, thank you for taking a look at this PR! I addressed your comment on the pairwise distance calculation, but I do not see the comment on scoring function evaluation. Did you make the comment in the tests or the notebook? (the scoring function is in both those places).

Sorry for the late reply @btjanaka , I did not have the time to post it yesterday, and I also was hesitating to add additional comments regarding the repertoire management (I think the calculation of the mode and of the spread should be done in the algorithm, and the repertoire addition condition should be receiving the same kinds of inputs as the MOMERepertoire). On further consideration, I will keep those comments for another PR, whose purpose will be to make the repertoire of MELS more modular.

[Egiob]

Hey @btjanaka, thanks for the very-clean PR, amazing work! I've left a couple of small suggestions, but all good for me 👍

[Egiob]

Minor suggestion: perhaps these functions belong elsewhere in a utils/ file ? I feel like they can be used by other methods as scores or metrics

[btjanaka]

I think if it becomes something used in multiple places, perhaps it can then be refactored? Since it is a private function, it seems it can be moved around pretty easily.

[btjanaka]

Thank you @Egiob! I've resolved all your comments except for the one about moving _dispersion

[btjanaka]

@Lookatator Is there anything else I should do to get this PR merged?

[manon-but-yes]

Hi @btjanaka thank you for this PR! Apologies for the delay I just got back to work.

[Lookatator]

@btjanaka I merged it into a local branch to check potential style issues before merging into develop :)