In R there are implementations of both the minimal CMA-ES code (also called
bareCMAES or pureCMAES, the Wikipedia version) and an adaption of the
Matlab code in parts. We have done lots of tests with these
implementations. It turned out in too many cases that the implementations
are unstable and may stop with errors. This is also described in a survey
article on global optimization to be published soon.
On the other hand, the full Matlab implementation is very useful and helped
me to solve some difficult real-world applications (where other
implementations stopped running). Therefore, I believe it would be
necessary and indeed worth to re-implement the procedure from the original
articles _and_ taking a look into the available implementations at times.
As Arnold Neumaier once wrote on
I found CMA-ES quite robust in dimensions up to 50.
(It gets very slow though when the dimension is large.)
On Tuesday, April 29, 2014 1:09:27 PM UTC+2, Nikolaus Hansen wrote:
>
> On Wednesday, 4 September 2013 16:29:28 UTC+2, Isaiah wrote:
>
> Maybe this is less useful, but the basic/demo version (by the author
>> himself) is public domain:
>> https://www.lri.fr/~hansen/barecmaes2.py
>>
>> And the Apache commons Java version claims to be Apache licensed, though
>> it is explicitly derived from the GPL matlab code.
>>
>
> the claim is correct, the Apache license has been granted by the author of
> the Matlab code.
>
