This is a TensorFlow based implementation for our paper on Exploration via Flow-Based Intrinsic Rewards.
Flow-based intrinsic module (FICM) is used for evaluating the novelty of observations. FICM generates intrinsic rewards based on the prediction errors of optical flow estimation since the rapid change part in consecutive frames usually serve as important signals.
Without any external reward, FICM can help RL agent to play SuperMario successfully.
This repo is inherited from large-scale-curiosity, and we also borrowed correlation layer
from flownet2_tf.
- Python3.5
pip install -r requirement.txt
pip install git+https://github.com/openai/baselines.git@3301089b48c42b87b396e246ea3f56fa4bfc9678
If you want to use FICM-C architecture, you'll need to compile for correlation layer additionally.
cd correlation_layer
make all
Note: You might encounter some errors raised from the source codes of tensorflow, you can easily change the header file of 'cuda_kernel_helper.h', 'cuda_device_functions.h', and 'cuda_launch_config.h'
If you want to train an agent to play SuperMario, you need to install retro
and import SuperMarioBros-Nes
game.
pip install retro
Read the official guide here
python run.py --feat_learning flowS --env_kind mario
python run.py --feat_learning flowS --env SeaquestNoFrameskip-v4 --seed 666
@article{flowbasedcuriosity2019,
Author = {Hsuan-Kung Yang, Po-Han Chiang, Min-Fong Hong and Chun-Yi Lee.},
Title = {Exploration via Flow-Based Intrinsic Rewards},
Booktitle = {arXiv:1905.10071},
Year = {2019}
}
@inproceedings{largeScaleCuriosity2018,
Author = {Burda, Yuri and Edwards, Harri and
Pathak, Deepak and Storkey, Amos and
Darrell, Trevor and Efros, Alexei A.},
Title = {Large-Scale Study of Curiosity-Driven Learning},
Booktitle = {arXiv:1808.04355},
Year = {2018}
}