How to train on low-dimensional Gym envs? about dreamer HOT 1 CLOSED

Thanks for your question. We have run PlaNet on low-dim inputs in the past and it worked well. The changes we made were to swap the encoder and decoder from conv layers to dense layers and to train more frequently. I think the same would work for Dreamer:

• Change ConvEncoder in models.py to look at the proprioceptive keys instead and use dense layers, e.g. 4 hidden layers of size 400.
• Change ConvDecoder in models.py in a similar way to predict a distribution over low-dim inputs.
• Change this line in dreamer.py to compute the log-prob of the low-dim data instead of the image.
• Comment out the image summaries because we're not predicting images anymore.

The code is already quite fast so I would try this first. When it works, you can also disable rendering in the environment to further speed things up and save RAM in the replay buffer. You can also try changing --train_every to, say, 200.

from dreamer.