Evosax - OpenES

[RobertTLange]

• Reference: Salimans et al. (2017)
• evosax Source Code: https://github.com/RobertTLange/evosax/blob/main/evosax/strategies/open_es.py
• By default the strategy will reshape the fitness using centered ranks.
• Logs and configurations: https://github.com/RobertTLange/evojax-benchmarks
• Benchmarks and hyperparameters:

	Benchmarks	Parameters	Results (Avg)
CartPole (easy)	900 (max_iter=1000)	Link	929.4153
CartPole (hard)	600 (max_iter=1000)	Link	604.6940
MNIST	90.0 (max_iter=2000)	Link	0.9669
Brax Ant	3000 (max_iter=1200)	Link	6726.2100
Waterworld	6 (max_iter=500)	-	-
Waterworld (MA)	2 (max_iter=2000)	-	-

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. I am currently running a brax grid search where I used a population size of 256 but with 500 iterations. I will add the results once they are done.

Cartpole-Easy	Cartpole-Hard	MNIST	Brax

Update: Added the brax-ant gridsearch. Very interesting to see that the hyperparameter ranges appear to be fairly task sensitive. The harder brax task appears to be less robust. Also interestingly the same qualitative patterns appeared in the ARS grid search (note: this used a different range).

[alantian]

Thanks, it's merged after some fixed for delayed import.

[RobertTLange]

Thank you. No worries. Given the current state of the world. It is completely understandable.