search

google / evojax

Apache License 2.0
834 stars 85 forks source link

Added simple_ga.py algo file #5

Closed MaximilienLC closed 2 years ago

lerrytang commented 2 years ago

Hi, thanks for the PR, I'm testing its performance on the tasks. At the same time, can you take a look at my review comment and make changes accordingly? Please see our implementation as an example.

lerrytang commented 2 years ago

Test results

Benchmarks Parameters Results
MNIST 90.0 (max_iter=5000) sigma=0.001 90.8
CartPole (easy) 900 (max_iter=2000) default 925
CartPole (hard) 600 (max_iter=2000) default 616
Waterworld 6 (max_iter=2000) default 6.39
Waterworld (MA) 2 (max_iter=5000) default 1.19
Brax Ant 3000 (max_iter=1000) truncation_divisor= 4 2336

Notes

  1. This table shows the implementation's performance so that users can choose the algorithms for their experiments.
  2. We are aware that some algorithms have limitations (e.g., unable to train large policy network), the benchmarks are therefore not hard requirements. However, we refuse to merge if some scores are significantly lower.
  3. After we release the test scripts, the PR submitter will be responsible for producing this table. We can help fill in some entries if the submitter cannot run the experiments due to hardware limitations.