MattStammers/appo-atari_freeway-approaching_sota

A(n) ASAC model trained on the atari_freeway environment.

This model was trained using Sample-Factory 2.0: https://github.com/alex-petrenko/sample-factory. Documentation for how to use Sample-Factory can be found at https://www.samplefactory.dev/

Downloading the model

After installing Sample-Factory, download the model with:

python -m sample_factory.huggingface.load_from_hub -r MattStammers/asac-atari_freeway

Using the model

To run the model after download, use the enjoy script corresponding to this environment:

python -m sf_examples.atari.enjoy_atari --algo=ASAC --env=atari_freeway --train_dir=./train_dir --experiment=asac-atari_freeway

You can also upload models to the Hugging Face Hub using the same script with the --push_to_hub flag. See https://www.samplefactory.dev/10-huggingface/huggingface/ for more details

Training with this model

To continue training with this model, use the train script corresponding to this environment:

python -m sf_examples.atari.train_atari --algo=ASAC --env=atari_freeway --train_dir=./train_dir --experiment=asac-atari_freeway --restart_behavior=resume --train_for_env_steps=10000000000

Note, you may have to adjust --train_for_env_steps to a suitably high number as the experiment will resume at the number of steps it concluded at.

SOTA Performance

This model is nearing SOTA performance for the Freeway environment: https://www.endtoend.ai/envs/gym/atari/freeway/ beating TQC and certainly DQN/PPO who both failed to converge after 10 million timesteps.

The composite score at 10 million timesteps is ~32 which is only two points off SOTA of 34. It appears that with PPO even after 2BN timesteps performance can only reach 33.6 - https://huggingface.co/edbeeching/atari_2B_atari_freeway_3333

I suspect that as with QR-DQN the SAC and TQC models can reach 34 - they just need more training to do so.