Module policy 

Policy and neural network implementations inference submodule available for WASM, training modules require training feature Policy and neural network wrappers.

After phase 5 of the Burn migration (#82), all policy networks live on the Burn backend. The pure-Rust WASM inference path (inference and universal_inference modules) is independent of Burn and is available without the training feature.

Modules

atari_cnn

Nature-DQN-scale CNN policies (actor-critic + Q-network) for the Atari (ALE) workload. Burn-backend Nature-DQN-scale CNN policy for the Atari (ALE) workload (Epic #306, Phase 3 — issue #327).

continuous_q

Continuous-action Q(s, a) critic for SAC (twin critics + targets), with hard and Polyak (soft) target-sync helpers. Burn-backend continuous-action Q-critic for SAC (part of #136).

inference

Inference-only model format for WASM deployment

lstm

LSTM (recurrent) actor-critic policy — Phase 1 of the recurrent-policy epic (#262). Carries memory across timesteps via a Burn 0.21 Lstm trunk, with forward_step for rollout collection and evaluate_sequences for the rank-3 training forward. Burn-backend LSTM (recurrent) actor-critic policy.

mlp

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

multi_discrete_mlp

Multi-discrete MLP policy used by Bucket Brigade and similar multi-discrete action spaces. Burn-backend multi-discrete actor-critic MLP for factored action spaces.

q_network

DQN Q-network with the same MLP backbone as MlpPolicy but with a single Q-head. Burn-backend Q-Network for DQN training (phase 4 of the Burn migration, #65).

sac_actor

SAC stochastic Gaussian actor (tanh-squashed) for continuous control. Burn-backend stochastic Gaussian actor for SAC (Soft Actor-Critic).

seeded_init

Seeded, host-side weight-initialization helpers that make policy construction bit-exact under PsroConfig::seed / NfspConfig::seed (issue #135). Burn 0.21’s Initializer has no seed parameter, so we pre-compute the trunk + head weights from an StdRng instead. Seeded, host-side weight-initialization helpers for bit-exact policy construction.

snake_cnn

3-conv + 2-fc CNN used by the Snake trainer. Burn-backend CNN policy for the Snake environment (phase 4 of the Burn migration, #65).

universal_inference

Universal inference system for diverse neural network architectures