world-models

Implementation of https://arxiv.org/abs/1803.10122

Requirments

keras
tensorflow
numpy
pycma

Training Process

1. collect_data.py
2. vae.py
3. collect_z.py
4. rnn.py
5. train_env.py

The controller is currently configured for training with the V model only (only z as input, no h).

Reward over episodes during training, with population of 12, and 4 trials per episode:

Environment used (removed zoom and indicators):
https://github.com/justinledford/gym/tree/car-remove-zoom

Data / Models

Data: https://s2.smu.edu/~jledford/ml/world-models/data.tar.gz
Models: https://s2.smu.edu/~jledford/ml/world-models/models.tar.gz

References

https://worldmodels.github.io/
https://blog.keras.io/building-autoencoders-in-keras.html
http://blog.otoro.net/2015/11/24/mixture-density-networks-with-tensorflow/
https://medium.com/applied-data-science/how-to-build-your-own-world-model-using-python-and-keras-64fb388ba459