feat: uncouple hiddenstate form previous layer in recurrent policies

instadeepai / Mava

🦁 A research-friendly codebase for fast experimentation of multi-agent reinforcement learning in JAX

Apache License 2.0

709 stars 83 forks source link

feat: uncouple hiddenstate form previous layer in recurrent policies #1059

Closed RuanJohn closed 6 months ago

RuanJohn commented 6 months ago

What?

The size of the agent hidden states are current coupled to the dimension of the preceding layer. This PR uncouples and allows for setting the hidden state sizes in the network config.

Extra

Maintains the current defaults. Ie. the hidden state size is 128 since the size of the last layer in the pre torso is 128.