Implementation of IPOP CMAES

[SuvarshaChennareddy]

This pull request is an initial implementation of IPOP CMAES, which is part of my Google Summer of Code (GSoC) project. Tests have been added, but not all of them are passing.

[SuvarshaChennareddy]

@zoq What should we do about the tests? Keep fine tuning the hyperparameters?

[zoq]

@zoq What should we do about the tests? Keep fine tuning the hyperparameters?

I'm looking into it a little bit more, in the sense of, checking if this is not caused by a hidden bug.

[SuvarshaChennareddy]

@zoq Using the Rastrigin Function now for testing, should (hopefully) be more stable.

[zoq]

Looks good, let's remove the Schwefel function.

[SuvarshaChennareddy]

Looks good, let's remove the Schwefel function.

Yup, done👍. Fixed the style as well.

[zoq]

Looks like the newly added functions are unstable.

[SuvarshaChennareddy]

Looks like the newly added functions are unstable.

Weird...the tests seemed stable on my system.

[zoq]

Did you test it with different random seeds? ensmallen_tests "[IPOPCMAESTest]" -rng-seed 'time'

[SuvarshaChennareddy]

Did you test it with different random seeds? ensmallen_tests "[IPOPCMAESTest]" -rng-seed 'time'

Yes

[zoq]

@mlpack-jenkins test this please

[SuvarshaChennareddy]

@zoq I'm sorry, but I'm currently occupied with other responsibilities. However, I'll certainly revisit this when I have more availability.

[SuvarshaChennareddy]

@zoq I ran a batch script with the following: FOR /L %%i IN (1,1,100) DO ( ensmallen_tests.exe "[IPOPCMAESTest]" --durations yes --rng-seed "time" ) I got 6 fails (out of a 100 tests) However, this doesn't seem to be reflected in the ensmallen CPU test. I don't really know how else I could test this.

Edit: Never mind, I was testing with PR #375's implementation of IPOP-CMA-ES, which includes the 'saveState' feature.

[SuvarshaChennareddy]

Further modifications to the CMA-ES and Active-CMA-ES implementations will likely be necessary because the optimizer is unable to converge, possibly due to the diverging step size.

[SuvarshaChennareddy]

@mlpack-jenkins test this please

[SuvarshaChennareddy]

@mlpack-jenkins test this please

[SuvarshaChennareddy]

A test of Demon SGD keeps failing. I'm going to try once again. I hope I'm not being a bother 💀 .

[SuvarshaChennareddy]

@mlpack-jenkins test this please

[SuvarshaChennareddy]

@mlpack-jenkins test this please

[SuvarshaChennareddy]

@zoq @rcurtin (sorry for the mention) I'm not sure what to do. I fixed some more inconsistencies (hopefully all of them) I found in the CMA-ES implementations. However, there is one test of Demon SGD that keeps failing consistently (just this one test). I have run the CMA-ES tests locally and they seem to be stable on my machine (a LOT more stable than before, probably because most of the major inconsistencies, if not all, have been fixed). Unfortunately, I don't think I can do anything else. I'd like to hear your thoughts.

[zoq]

Happy to look into it.

[mlpack-bot[bot]]

This issue has been automatically marked as stale because it has not had any recent activity. It will be closed in 7 days if no further activity occurs. Thank you for your contributions! :+1:

[mlpack-bot[bot]]

This issue has been automatically marked as stale because it has not had any recent activity. It will be closed in 7 days if no further activity occurs. Thank you for your contributions! :+1:

[coatless]

@zoq did you have a moment to check on what @SuvarshaChennareddy mentioned?

[zoq]

@zoq did you have a moment to check on what @SuvarshaChennareddy mentioned?

yes, will check