This repository contains benchmark results, helper scripts, ES configurations and logs for testing the performance of evolutionary strategies in evojax. These can come in handy, when aiming to merge a new JAX-based ES into the projects.
[OPTIONAL] Tune hyperparameters using mle-hyperopt. Here is an example for running a grid search for ARS over different learning rates and perturbation standard deviations via:
This will sequentially execute 25 ARS-MNIST evolution runs for a grid of different learning rates and standard deviations. After the search has completed, you can access the search log at log/ARS/mnist/search_log.yaml
Note: For the brax environment I reduced the population size from 1024 to 256 and increased the search iterations by the same factor (300 to 1200) in the main run. For the grid search I used a population size of 256 but with 500 iterations.