README.md

World Models A3C

Implementation of a variant of World Models

Note

• Replaced MDN-RNN to LSTM for Memory
• Replaced CMA-ES to A3C for Controller
• Trained over two stages
  • Stage 1: V and M were trained on dataset with random rollout
  • Stage 2: V and M were trained on dataset with a3c rollout

Training Result

Result with dataset using random rollout

Result with dataset using the pretrained model rollout

Play Demo

---

Environment Setting

apt-get update
apt-get install swig
pip install gym[box2d]

Training Stage I

Dataset Generation using Rollout with random policy

python rollout.py

Vision model with VAE

python train-vae.py

Memory model with LSTM-RNN

python train-rnn.py

Controller with A3C

python train-a3c.py

Training Stage II

Rollout with the pretrained model

python rollout-a3c.py

Fine-tuning V and M with new dataset

vi hparams.py
    extra = True

python train-vae.py
python train-rnn.py

Train new C with the improved V and M

python train-a3c.py

Test

# <# of plays> <seed> <is_record>
python test.py 2 999 False

Reference

• https://arxiv.org/abs/1803.10122
• https://github.com/ctallec/world-models