MAP-CMA (PPSN2024) #186
MAP-CMA (PPSN2024) #186
Conversation
|
I have added the comment, "MAP-CMA does not employ the Heaviside function h_sigma for simplifying the update rules." to _mapcma.py. |
| @@ -317,6 +317,53 @@ The full source code is available [here](./examples/catcma.py). | |||
| </details> | |||
|
|
|||
|
|
|||
| #### Maximum a Posteriori CMA-ES [Hamano et al. 2024b] | |||
| MAP-CMA is a method that is introduced to interpret the rank-one update in the CMA-ES from the perspective of the natural gradient. | |||
| The rank-one update derived from the natural gradient perspective is extensible, and an additional term, called momentum update, appears in the update of the mean vector. | |||
There was a problem hiding this comment.
It is better to mention that like "The performance of MAP-CMA is not so different from CMA-ES, as the main motivation of MAP-CMA comes from theoretical understanding of CMA-ES".
README.md
Outdated
|
|
||
| if __name__ == "__main__": | ||
| dim = 20 | ||
| optimizer = MAPCMA(mean=3 * np.ones(dim), sigma=2.0, momentum_r=dim) |
There was a problem hiding this comment.
It's strange that the optimization on the Rosenbrock function starts from m=3. Please adjust it to m=0.0 and sigma=0.5. The sigma setting is to avoid local optima and make the check easier.
cmaes/_mapcma.py
Outdated
|
|
||
| mu = population_size // 2 | ||
|
|
||
| # MAPCMA uses positive weights |
There was a problem hiding this comment.
The reason for this is not clear. Please add a brief explanation if possible.
|
|
||
|
|
||
| class MAPCMA: | ||
| """MAP-CMA stochastic optimizer class with ask-and-tell interface. |
There was a problem hiding this comment.
It is helpful to mention the difference from the CMA-ES to understand MAP-CMA.
Please add an explanation like "The only difference from the CMA-ES is the mean vector update..."
|
@ha-mano I left some comments. Can you check them? |
|
@nomuramasahir0 Thank you for your review. I have addressed the comments and made the necessary changes. |
|
Thanks! LGTM. |
I have implemented MAP-CMA [Hamano+, PPSN2024] and created the description for the README.
Note that in the README, the abbreviation of the reference [Hamano et al. 2024] has been changed to distinguish between GECCO2024 and PPSN2024.