Module ppo 

PPO trainer (Burn backend).

After phase 5 of the Burn migration (#82), Burn is the only tensor backend in the workspace. The historical train::ppo_burn parallel sibling has been collapsed back into train::ppo. The trainer struct retains its PPOTrainerBurn<B, P, O> name because Burn’s optimizer-consumes-module ownership model is structurally different from a hypothetical in-place trainer.

Contents

• config — PPOConfig hyperparameters / builder API.
• stats — TrainingStats / AggregatedStats per-update metrics.
• loss — backend-generic PPO loss math (policy/value/entropy) and the generate_minibatch_indices helper.
• trainer — PPOTrainerBurn<B, P, O> that owns the policy module (Burn’s optimizer-consumes-module ownership model) and exposes a train_step that runs the surrogate-loss / gradient-step / KL early stop logic.

Re-exports

pub use config::PPOConfig;

pub use loss::compute_entropy_loss;

pub use loss::compute_policy_loss;

pub use loss::compute_value_loss;

pub use loss::generate_minibatch_indices;

pub use loss::scalar_f64;

pub use stats::AggregatedStats;

pub use stats::TrainingStats;

pub use trainer::PPOTrainerBurn;

Modules

config

PPO configuration and hyperparameters

loss

PPO loss math (Burn backend).

stats

Training statistics for PPO

trainer

Burn-backend PPO trainer (phase 3 of the Burn migration, #80).