Paper link: arXiv

Code for paper Efficient Reinforcement Learning in Block MDPs: A Model-free Representation Learning approach

Creating a virtual environment is recommended (or using conda alternatively):

pip install virtualenv
virtualenv /path/to/venv --python=python3

#To activate a virtualenv: 

. /path/to/venv/bin/activate

To install the dependencies (the results from the paper are obtain from gym==0.14.0):

pip install -r requirements.txt

To install pytorch, please follow PyTorch. Note that the current implementation does not require pytorch gpu.

We use wandb to perform result collection, please setup wandb before running the code or add os.environ['WANDB_MODE'] = 'offline' in main.py.

To reproduce our result in comblock (Section 6.1), please run:

bash run.sh 

For online reptransfer, please run:

bash run_online.sh 

To reproduce our result in comblock with partitioned observation (Section 6.2), please run:

bash run_po.sh 

For online reptransfer, please run:

bash run_po_online.sh

To see all the hyperparameters, please refer to utils.py.