Module loss 

PPO loss math (Burn backend).

After phase 5 of the Burn migration (#82), Burn is the only tensor backend in the workspace; this module hosts the canonical PPO loss surface used by crate::train::ppo::trainer::PPOTrainerBurn.

API notes

• Burn’s mean() returns a rank-1 scalar tensor — there is no Kind::Float argument, and the resulting tensor still carries the backend parameter. We expose the scalar-extracting helper (scalar_f64) that the trainer needs at the loss boundary.
• Burn carries the const rank in the type, so the policy-loss / value-loss tensors carry their concrete rank (Tensor<B, 1> for per-row, scalar at the end).
• We extract scalars via scalar_f64 which detaches implicitly via .into_scalar(); no explicit no_grad scope is needed because the autograd tape is not extended beyond the loss tensors we return.

Functions

compute_entropy_loss

Entropy loss = -mean(entropy) (we minimize loss, so we negate the entropy bonus before adding it into the total loss).

compute_policy_loss

Compute PPO policy (surrogate) loss with clipping.

compute_value_loss

Compute the value-function loss with optional clipping.

generate_minibatch_indices

Generate randomized minibatch indices from a flat rollout buffer.

generate_minibatch_indices_with_rng

Generate randomized minibatch indices from a flat rollout buffer using a caller-supplied RNG.

scalar_f64

Read a scalar tensor’s value into an f64.