Creating Classical Evolution Strategies

Hello Everyone,

I am working on implementing Classical Evolution Strategies in Ray. I am new to Ray, and I am struggling trying to find the best way to create an N sized population where every individual has the same neural network but different weights.

My goal is to create an algorithm similar to the included Population Based Training(PBT), but focused on weights instead of hyperparameters, where I can personalize the selection, noise and resampling process.

Any guidance or tips would be greatly appreciated.
Thanks in advance!

PS: I have seen Ray’s ES algorithm but its implementation is completely different and it only generates 1 model

During the RL Summit end March, Siemens talked about using Offline, Population-based, gradient-free, not BC approach.

Here is link to their talk, slides, and paper. Not sure if it is in right direction what you’re looking for.