/liberty

Primary LanguagePythonMIT LicenseMIT

LIBERTY:Efficient Potential-Based Exploration in Reinforcement Learning Using the Inverse Dynamic Bisimulation Metric

This is the official implementation of NeurIPS 2023 paper - [Efficient Potential-Based Exploration in Reinforcement Learning Using the Inverse Dynamic Bisimulation Metric]

illustrations illustrations

Requirements

  • cuda-10.0
  • pytorch==1.2.0
  • gym==0.12.5
  • gym[atari]
  • gym_super_mario_bros
  • mujoco-py==1.50.1.56
  • mpi4py
  • pillow
  • tqdm

Installation

  1. Please install the required packages in the above list.
  2. Install utils:
pip install -e .

Run Experiments

Please enter ppo folder to run LIBERTY

cd ppo/

Instructions

python -u train.py --env-name='HalfCheetah-v2' --cuda (if cuda is available) --log-dir='logs' --seed=777