Introduction

This repo contains code for learning reward functions and state weights from preferences. After installation, find the task to run in main.py. For example,

python -um crowd_pbrl.main --exp_path <path-to-experiments> --gpu_id 0 --split 1 --game BeamRider --task sample_trajectories --method bt

Samples some trajectories for reward learning. Make sure there is a config.gin file in your . An example for configuration file is provided as config.gin. A script for how one can run experiments on Enduro is provided as run.py.

Installation

1. (On windows with 1080Ti) Setup a virtual Python 3.7 environment and install other dependencies. If you are using miniconda,

conda create -n crowd_pbrl python==3.7
conda activate crowd_pbrl
conda install cudatoolkit=10.0
conda install cudnn
pip install tensorflow-gpu==1.15
pip install dopamine-rl==3.0.1
pip install ffmpeg-python
pip install scikit-learn
pip install gym[atari]
pip install matplotlib

copy atari files

python -m atari_py.import_roms <ROM_path>