Dream to Control

Fast and simple implementation of the Dreamer agent in TensorFlow 2.

If you find this code useful, please reference in your paper:

@article{hafner2019dreamer,
  title={Dream to Control: Learning Behaviors by Latent Imagination},
  author={Hafner, Danijar and Lillicrap, Timothy and Ba, Jimmy and Norouzi, Mohammad},
  journal={arXiv preprint arXiv:1912.01603},
  year={2019}
}

Method

Dreamer learns a world model that predicts ahead in a compact feature space. From imagined feature sequences, it learns a policy and state-value function. The value gradients are backpropagated through the multi-step predictions to efficiently learn a long-horizon policy.

Instructions

Get dependencies:

pip3 install --user tensorflow-gpu==2.1.0
pip3 install --user tensorflow_probability
pip3 install --user git+git://github.com/deepmind/dm_control.git
pip3 install --user pandas
pip3 install --user matplotlib

Train the agent:

python3 dreamer.py --logdir ./logdir/dmc_walker_walk/dreamer/1 --task dmc_walker_walk

Generate plots:

python3 plotting.py --indir ./logdir --outdir ./plots --xaxis step --yaxis test/return --bins 3e4

Graphs and GIFs:

tensorboard --logdir ./logdir

alec-tschantz/dreamer

Dream to Control

Method

Instructions