Default Model Size Question

RLlib
1

Hi,

Just a quick question, when using RRLIB algorithms, when no model is specified it defaults to a fcnet (fully-connected net?). I just wanted to understand correctly that that model then makes 256 layers with each having input and output size of observation_space.shape?

Thanks!

2

Hi,

yes, that’s correct. The default model for standard environments is a fully connected MLP with two hidden layers and 256 hidden units per layer. The size of the input and output layer are automatically adjusted to the size of your observations and actions.

For environments with 2D inputs (e.g., images), RLlib uses a different model by default with convolutions.

Check the documentation here for details: RLlib Models, Preprocessors, and Action Distributions — Ray v1.2.0

Does this answer your question?

3 Likes
3

Yup, that answers my question, thank you!

2 Likes