Skip to content

[Proposal] Support 1D search space in CMA-ES#190

Merged
nomuramasahir0 merged 1 commit intoCyberAgentAILab:mainfrom
y0z:feature/support-1d
Nov 12, 2024
Merged

[Proposal] Support 1D search space in CMA-ES#190
nomuramasahir0 merged 1 commit intoCyberAgentAILab:mainfrom
y0z:feature/support-1d

Conversation

@y0z
Copy link
Copy Markdown
Contributor

@y0z y0z commented Oct 31, 2024 •
edited
Loading

Motivation

At the moment, cmaes does not support 1D search space.
However, based on my understanding, the CMA-ES algorithm can naturally handle 1D search space, so I want to support it.
It has also been observed that some users want to support this (optuna/optuna#5524).

In this PR, I support 1D search space in CMA-ES.

Description of Changes

  • Change an assertion in _cma.py to ensure n_dim > 0 rather than n_dim > 1.

I confirmed that the code with the above change passes all tests and works fine in the following toy example.

import numpy as np
from cmaes import CMA

def quadratic(x1):
    return (x1 - 3) ** 2

if __name__ == "__main__":
    optimizer = CMA(mean=np.zeros(1), sigma=1.3)

    for generation in range(20):
        solutions = []
        for _ in range(optimizer.population_size):
            x = optimizer.ask()
            value = quadratic(x[0])
            solutions.append((x, value))
            print(f"#{generation} {value} (x1={x[0]})")
        optimizer.tell(solutions)

Output: It is confirmed that the search converges towards the optimum (x1 = 3).

#0 16.643669924292066 (x1=-1.0796654181797876)
#0 7.2004261524959166 (x1=0.3166390193460894)
#0 8.189839736380824 (x1=0.1382103962064533)
#0 12.497988920536582 (x1=-0.5352494849071942)
#1 7.179325689223134 (x1=0.3205736268329496)
#1 3.6146114323456273 (x1=1.0987868524687645)
#1 5.255710328044975 (x1=0.7074663954382319)
#1 10.095710257035286 (x1=-0.17737474293406053)
...
#18 6.963502924227799e-05 (x1=2.9916552394137232)
#18 0.002559389599251478 (x1=2.94940958984895)
#18 0.004124398005827681 (x1=3.0642214762040525)
#18 7.020158315938154e-05 (x1=2.991621361497273)
#19 0.0063421454828861525 (x1=3.0796375883793963)
#19 6.989768552220104e-05 (x1=2.991639516430122)
#19 0.0026858233887192963 (x1=2.948175069814622)
#19 0.0012593115389895819 (x1=3.0354867797776803)

@nomuramasahir0 nomuramasahir0 self-requested a review November 12, 2024 02:36
@nomuramasahir0
Copy link
Copy Markdown
Contributor

nomuramasahir0 commented Nov 12, 2024

Thank you for the PR! As you mentioned, CMA-ES can be applied to 1-D problems (in this case, CMA-ES essentially reduces to ES), although I haven't empirically tested it yet. Some hyperparameters, like the learning rate, might need adjustment, so let's see how it performs.

@nomuramasahir0 nomuramasahir0 merged commit 291376e into CyberAgentAILab:main Nov 12, 2024
@y0z y0z deleted the feature/support-1d branch November 12, 2024 04:43
@sochinbi sochinbi mentioned this pull request Dec 4, 2024
Closed
@sawa3030 sawa3030 mentioned this pull request Aug 1, 2025
Merged
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@nomuramasahir0 nomuramasahir0 Awaiting requested review from nomuramasahir0

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

2 participants

@y0z @nomuramasahir0