Using State Predictions for Value Regularization in Curiosity Driven Deep Reinforcement Learning

About

Tensorflow implementation of the algorithm described in ‘Using State Predictions for Value Regularization in Curiosity Driven Deep Reinforcement Learning’ using the maze environments.

Installation

sudo apt-get install -y tmux htop cmake golang libjpeg-dev
git clone -b master https://github.com/ManuelFritsche/vpc.git
conda create -n curiosity python=2.7
source activate curiosity
pip install numpy
pip install -r vpc/requirements.txt

Training

cd vpc/
# for A3C remove --unsup, for PRED use --unsup pred, for VPC use --unsup vpc
# for Maze A use --env-id mazeSmall-v0, for Maze B use --env-id mazeLarge-v0
python train.py --unsup vpc --env-id mazeSmall-v0

Training process is shown in Tensorboard on http://localhost:12345

Acknowledgements

The implentation is based on the code of Curiosity-driven Exploration by Self-supervised Prediction.
Vanilla A3C code is based on the open source implementation of universe-starter-agent.
Maze implementations are based on Pycolab