Skip to content
Code for human intervention reinforcement learning
Branch: master
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
docker
humanrl
models
scripts
train
universe-starter-agent
.atomignore
.dockerignore
.gitignore
DOCKER-INSTRUCTIONS.md
LICENSE
README.md
base.docker
build_docker.sh
build_docker_gpu.sh
environment.yml
mac_run_docker.sh
main.docker
setup_conda.sh

README.md

Human intervention reinforcement learning

Research code for the paper "Trial without Error: Towards Safe Reinforcement Learning via Human Intervention" (arxiv) (2017)

Contributors (alphabetical): Owain Evans, Vlad Firoiu, Girish Sastry, William Saunders

Overview

This repository contains the code for human intervention reinforcement learning in Atari environments (based on OpenAI's Gym). The humanrl package contains various Gym environment wrappers and utilities that allow modifying Atari environments to include catastrophes.

scripts/human_feedback.py is a script that allows a human to intervene during offline or online training of an RL agent.

Installation and use

To label and run the code locally, first create an Anaconda environment with our packages:

conda env create
source activate humanrl

See the human feedback README for directions on providing human feedback with the OpenAI universe starter agent.

See the catastrophe wrapper for a general purpose way to add catastrophes to Gym environments.

You can’t perform that action at this time.