RPSF: Recovering Permuted Sequential Features for Effective Reinforcement Learning

This is a PyTorch implementation of SVEA-C-RPSF and SVEA-T-RPSF using Convolution Neural Networks and Vision Transformers respectively.

Setup

We assume that you have access to a GPU with CUDA >=9.2 support. All dependencies can then be installed with the following commands:

cd ./cnn
conda env create -f ./setup/conda.yaml
conda activate svea-c-rpsf
sh ./setup/install_envs.sh

SVEA-C-RPSF and SVEA-T-RPSF use the same dependencies.

Training & Evaluation

In the cnn and transformer directories, scripts directories contain bash scripts for SVEA-C-RPSF and SVEA-T-RPSF, which can be run by sh /cnn/scripts/svea-c-rpsf.sh and sh /transformer/scripts/svea-t-rpsf.sh respectively.

Alternatively, you can call the python scripts directly, e.g. for training of SVEA-C-RPSF call

python3 cnn/src/train.py --seed 0 --algorithm svea --use_aux

to run SVEA-C-RPSF on the default task, walker_walk, and using the default hyperparameters.