Curiosity-driven Exploration by Self-supervised Prediction

In ICML 2017 [Project Website] [Demo Video]

Deepak Pathak, Pulkit Agrawal, Alexei A. Efros, Trevor Darrell
University of California, Berkeley

This is a tensorflow based implementation for our ICML 2017 paper on curiosity-driven exploration for reinforcement learning. Idea is to train agent with intrinsic curiosity-based motivation (ICM) when external rewards from environment are sparse. Surprisingly, you can use ICM even when there are no rewards available from the environment, in which case, agent learns to explore only out of curiosity: 'RL without rewards'. If you find this work useful in your research, please cite:

@inproceedings{pathakICMl17curiosity,
    Author = {Pathak, Deepak and Agrawal, Pulkit and
              Efros, Alexei A. and Darrell, Trevor},
    Title = {Curiosity-driven Exploration by Self-supervised Prediction},
    Booktitle = {International Conference on Machine Learning ({ICML})},
    Year = {2017}
}

1) Installation and Usage

1. This code is based on TensorFlow. To install, run these commands:

# you might not need many of these, e.g., fceux is only for mario
sudo apt-get install -y python-numpy python-dev cmake zlib1g-dev libjpeg-dev xvfb \
libav-tools xorg-dev python-opengl libboost-all-dev libsdl2-dev swig python3-dev \
python3-venv make golang libjpeg-turbo8-dev gcc wget unzip git fceux virtualenv \
tmux

# install the code
git clone -b master --single-branch https://github.com/pathak22/noreward-rl.git
cd noreward-rl/
virtualenv curiosity
source $PWD/curiosity/bin/activate
pip install numpy
pip install -r src/requirements.txt
python curiosity/src/go-vncdriver/build.py

# download models
bash models/download_models.sh

# setup customized doom environment
cd doomFiles/
# then follow commands in doomFiles/README.md

2. Running demo

cd noreward-rl/src/
python demo.py --ckpt ../models/doom/doom_ICM
python demo.py --env-id SuperMarioBros-1-1-v0 --ckpt ../models/mario/mario_ICM

3. Training code

cd noreward-rl/src/
# For Doom: doom or doomSparse or doomVerySparse
python train.py --default --env-id doom

# For Mario, change src/constants.py as follows:
# PREDICTION_BETA = 0.2
# ENTROPY_BETA = 0.0005
python train.py --default --env-id mario --noReward

xvfb-run -s "-screen 0 1400x900x24" bash  # only for remote desktops
# useful xvfb link: http://stackoverflow.com/a/30336424
python inference.py --default --env-id doom --record

2) Other helpful pointers

• Paper
• Project Website
• Demo Video
• Reddit Discussion
• Media Articles (New Scientist, MIT Tech Review and others)

3) Acknowledgement

Vanilla A3C code is based on the open source implementation of universe-starter-agent.