Module mlp 

MLP actor-critic policy used by the CartPole / Pong / SimpleBandit PPO trainers. Burn-backend MLP actor-critic policy.

Implements a 2/3-layer MLP actor-critic architecture with orthogonal initialization (PPO recipe — gain sqrt(2) on the trunk, 0.01 on the output heads).

Entry points

• MlpBurnPolicy::new — the simple scout-era constructor (random Kaiming init, 2 layers).
• MlpBurnConfig — builder-style configuration with orthogonal init, activation, and depth knobs; supports the encoder-tap helper that downstream regularizers want.

Why generic over B: Backend?

Burn’s idiomatic pattern is to make every Module generic over a Backend type parameter (CPU NdArray, GPU Wgpu/Cuda, autodiff-decorated variants, etc.). Production trainers can re-use the same modules with a different backend at the top of the binary without touching the policy code.

Structs

MlpBurnConfig

Configuration for MlpBurnPolicy architecture.

MlpBurnPolicy

Two- or three-layer MLP actor-critic for discrete action spaces, ported to Burn.

MlpBurnPolicyRecord

The record type for the module.

MlpBurnPolicyRecordItem

The record item type for the module.

Enums

BurnActivation

Activation function applied between hidden layers in MlpBurnPolicy (and its multi-discrete sibling).