You Only Evaluate Once (YOEO)

Prerequisite

mujoco200
conda

• Remember to add the mujoco directory to LD_LIBRARY_PATH environment variable.
• All the other dependencies will be handled in the following installation script via conda.

Install

git clone <anonymized for dual blind>
cd <cloned-dir>
conda create --file env.yaml --name YOEO
conda activate YOEO
# in the case of error during creation, use conda update commands:
# conda env update --file env.yaml
pip install git+https://github.com/rail-berkeley/d4rl@master#egg=d4rl

Run

Train Y

```
python -m YOEO.scripts.value_training --log_dir ./log/directory/you/want --seed {seed you want} --config_file ./experiments/Y.gin ./experiments/envs/{hopper,walker2d,etc}.gin --config_params env_id=\'{d4rl env id, like hopper-medium-replay-v0}\'
```

Train YOEO (and ablations)

• YOEO

  python -m YOEO.scripts.offline_rl --log_dir ./log/directory/you/want --seed 1 --config_file ./experiments/ValidQbeta.gin ./experiments/envs/{hopper,walker2d,etc}.gin --config_params env_id=\'{d4rl env id, like hopper-medium-replay-v0}\' Y_chkpt=\'./path/to/trained/Y/value.tf\' lambda={0.1 or 1.0}
  

• $Q^\beta$ without regularizations

  python -m YOEO.scripts.offline_rl --log_dir ./log/directory/you/want --seed 1 --config_file ./experiments/Qbeta.gin ./experiments/envs/{hopper,walker2d,etc}.gin --config_params env_id=\'{d4rl env id, like hopper-medium-replay-v0}\'
  

• For other ablation studies, you can just change the gin file, or you can provide the hyperparameters using --config_params arguments, then the provided arguments will overide the arguments in gin file.

Observe training statistics & results

• Use Tensorboard

  tensorboard --logdir <log directory>