roberttlange / code-and-blog Goto Github PK

Hi Robert:

Thanks for sharing your implementation of CMA-ES. I found that the "chi_n" parameter is not initialized together with the rest of the parameters here:

"chi_n" appears here, for example:

From a few other implementations online, I found that it was sometimes heuristically initialized to:

"chi_n": jnp.sqrt(n_dim) * (1.0-1.0/(4.0n_dim) + 1.0/(21.0n_dim**2)),

and that this worked pretty well in practice for my problems.

Hi Robert,

Thanks for the super helpful implementation guide for CMA-ES in JAX. I came across an inconsistency in the implementation about the variable y_w in your code.

Eq. 45 from Hansen's tutorial looks like this

Eq. 41 defines $\left<y\right>_w$ as follows:

which is "step of the distribution mean disregarding step-size σ". The pure-python implementation you link in your tutorial honours Eq. 41.

https://github.com/CyberAgentAILab/cmaes/blob/7508e8a7fea0102932a6d57f4a7be528cce306e4/cmaes/_cma.py#L336

However, it seems that y_w is divided by the step size in your implementation.

Am I missing something here? I'm not entirely sure how big of a difference this makes either.

Hi. Thank you for sharing your blog. I've added it to my feed bin, but it looks like the RSS feed is not working, I only see the first 3 posts from 2019.

Great blog!

Quick note -- I believe the function below does not actually depend on x, even though I think you want it to.

def vmap_relu_layer(params, x): """ vmap version of the ReLU layer """ return jit(vmap(relu_layer, in_axes=(None, 0), out_axes=0))