Struct CandleStochasticPolicy 

pub struct CandleStochasticPolicy { /* private fields */ }

Implementations

impl CandleStochasticPolicy

pub fn new( obs_dim: usize, act_dim: usize, hidden: usize, lr: f64, device: Device, seed: u64, ) -> Result<Self, NNError>

impl CandleStochasticPolicy

pub fn sac_actor_step( &mut self, obs: &TensorData, alpha: f32, critic: &CandleTwinQ, ) -> Result<TrainMetrics, NNError>

SAC actor gradient step with autograd flowing through the critic.

Takes concrete CandleTwinQ to preserve gradient flow from Q-values back to actor parameters via the reparameterized actions.

Trait Implementations

impl StochasticPolicy for CandleStochasticPolicy

fn sample_actions( &self, obs: &TensorData, ) -> Result<(TensorData, TensorData), NNError>

Sample actions with reparameterization trick. Returns (squashed_actions [batch, act_dim], log_probs [batch]).

fn deterministic_action(&self, obs: &TensorData) -> Result<TensorData, NNError>

Deterministic action (mean through squashing).

fn learning_rate(&self) -> f32

fn set_learning_rate(&mut self, lr: f32)

fn save(&self, path: &Path) -> Result<(), NNError>

fn load(&mut self, path: &Path) -> Result<(), NNError>

impl Send for CandleStochasticPolicy

impl Sync for CandleStochasticPolicy