Module loss 

Behavioral Cloning loss math (Burn backend).

Behavioral cloning is supervised learning: the policy’s action head is trained to reproduce expert action labels via cross-entropy. Unlike the RL losses (crate::train::a2c::loss, crate::train::ppo::loss) there is no advantage, no importance ratio, no entropy bonus, and no value term — just the negative log-likelihood of the expert action under the current policy.

The loss is discrete-only, matching MlpBurnPolicy’s categorical action head. We re-export PPO’s host-side scalar_f64 so callers can pull the whole BC loss surface from bc::loss without reaching into ppo::loss.

Re-exports

pub use crate::train::ppo::loss::scalar_f64;

Functions

compute_bc_loss

Compute the supervised behavioral-cloning cross-entropy loss.