Multi-View Reinforcement Learning

This repo includes the poster and experiments for the paper Multi-View Reinforcement Learning.

Requirements

Test with python3.6.1 (might work with later versions). Install all requirements by

pip install -r requirements.txt

Code structure

data

extract.py/sh: Sample the data used in model learning

models

vrnn.py: Define the rnn structure. vae.py: Define the vae structure.

the rest

config.py: Define the env.
env.py: Preprocess the environment used in the model learning process.
train.py: Train the multi-view model.
utils.py and ops.py: Define many functions to simplify the multi-view model training code.
wrappers.py: Define the transformer for each view.

How to run

Download the data from here, extract the file, and put the extracted folder in ./data.

You can then run

python train.py --model-dir checkpoint/model --data-dir data/record --view transposed --gpu 0

The output and training log is saved at checkpoint/model, the data is loaded from data/record, the view choice is transposed and the selected gpu is 0.

Citation

If you found it helpful, consider citing the following paper:

@article{li2019multi,
  title={Multi-View Reinforcement Learning},
  author={Li, Minne and Wu, Lisheng and Ammar, Haitham Bou and Wang, Jun},
  booktitle={Advances in Neural Information Processing Systems},
  year={2019}
}

mlii/mvrl