Evolution strategies - make reproducible

Hi all!

I am running some experiments with the Evolution Strategies algorithm at RLlib, and I need to make the results reproducible. However, setting the “seed” in the configuration file as posted here (Reproducible training - setting seeds for all workers / environments) is not working. Do you know how I can have identical runs?

The configuration file that I use is:

env: Humanoid-v2
run: ES
episode_reward_mean: 200
# Works for both torch and tf.
framework: tf
num_workers: 2
seed: 42
train_batch_size: 10000


Hey @Fot , ah, I think this makes sense. RLlib does all the seed handling in the RolloutWorkers, which ES and ARS don’t use :confused: (one more good reason to force-move these two into using standard Policies/RolloutWorkers/etc…).

There is a PR here that fixes the issue: