PyTorch AWR

This repository contains a PyTorch implementation of the reinforcement learning algorithm Advantage Weighted Regression (AWR). The objective of this implementation is to make it possible for PyTorch users to use AWR for their RL projects, as the original implementation is for TensorFlow (see references).

Setup

• make sure you are running Python 3.6.9 or above
• run pip3 install -r requirements.txt --no-cache-dir (the no-cache-dir-option is sometimes required to finish the download of torch)
• you can remove mujoco-py from the requirements if you do not have a license
• edit main.py to configure your environment and hyper-parameters (cli options are planned)
• run pyton3 main.py

Features

• full implementation of AWR according to the paper
• hyper-parameters pre-filled with appropriate values
• training and testing framework: given the NN models, the environment and the hyper-parameters, the framework trains the models and conducts a series of tests on them after completion

References

The authors' code (written for TensorFlow) for the paper can be found here.