Synthetic Gymnax

💡 Make a one-line change ...

import gymnax
env, params = gymnax.make("CartPole-v1")

...  # your training code

import gymnax, synthetic_gymnax
env, params = gymnax.make("Synthetic-CartPole-v1")
# add 'synthetic' to env:  ^^^^^^^^^^
...  # your training code

💨 ... and enjoy fast training.

The synthetic environments are meta-learned to train agents within 10k time steps. This can be much faster than training in the real environment, even when using tuned hyperparameters!

• 🟩 Real environment training, using tuned hyperparameters (IQM of 5 training runs)
• 🟦 Synthetic environment training, using any reasonable hyperparameters (IQM performance of 20 training runs with random HP configurations)

🏅 Performance of agents after training for 10k synthetic steps

💫Replicating our results

We provide the configurations used in meta-training the checkpoints for synthetic environments in synthetic_gymnax/checkpoints/*environment*/config.yaml. They can be used with the meta-learning script by calling e.g.

python examples/metalearn_synthenv.py --config synthetic_gymnax/checkpoints/hopper/config.yaml

Please note that when installing via pip, the configs are not bundled with the package. Please clone the repository to get them.

✍ Citing and more information

If you use the provided synthetic environments in your work, please cite us as

@article{
  ...
}