pulsatrix::TrajectoryBalanceLoss Class Reference

Δ(τ) = log Zθ + Σ log P_F(s_{t+1}|s_t) − log R(x) − Σ log P_B(s_t|s_{t+1}), loss = Δ(τ)². More...

#include <trajectory_balance_loss.hpp>

Public Member Functions
	TrajectoryBalanceLoss ()=default
float	forward (float sum_log_pf, float sum_log_pb, float log_reward, float log_z)
	Computes Δ(τ) and the loss Δ(τ)², caching Δ for the accessors below.
float	delta () const
	Δ(τ) from the most recent forward() call.
float	grad_log_z () const
	d(loss)/d(log Z) = 2Δ – feed directly into LearnableScalar::accumulate_grad.
float	grad_weight_for_log_pf () const
	The weight w = -2Δ such that the per-trajectory-step policy-network gradient is w · (p[k] − 1{k==a_t}) – PolicyGradientLoss's exact softmax/log gradient identity shape, with this trajectory's Δ playing the role of the per-step return.

Detailed Description

Δ(τ) = log Zθ + Σ log P_F(s_{t+1}|s_t) − log R(x) − Σ log P_B(s_t|s_{t+1}), loss = Δ(τ)².

Note

Deliberately does not follow MSELoss/PolicyGradientLoss's single-tensor backward() shape: Δ feeds two heterogeneous consumers (a bare scalar log Z, and a sequence of per-step policy-network logits from different forward-policy calls at different states) that no single gradient tensor covers. Instead exposes delta() (the one real error signal) plus two named derived accessors. See mission_trajectory_balance_loss.md's Design section for the full derivation.

Not a Module subclass, for MSELoss's own reason: a loss is the seed point relevance propagation starts from, not something propagate_relevance is defined for.

Constructor & Destructor Documentation

TrajectoryBalanceLoss()

pulsatrix::TrajectoryBalanceLoss::TrajectoryBalanceLoss ( )

default

Member Function Documentation

delta()

float pulsatrix::TrajectoryBalanceLoss::delta

(

)

const

Δ(τ) from the most recent forward() call.

forward()

float pulsatrix::TrajectoryBalanceLoss::forward	(	float	sum_log_pf,
		float	sum_log_pb,
		float	log_reward,
		float	log_z
	)

Computes Δ(τ) and the loss Δ(τ)², caching Δ for the accessors below.

Parameters

sum_log_pf	Σ_t log P_F(s_{t+1}|s_t) over the whole trajectory.
sum_log_pb	Σ_t log P_B(s_t|s_{t+1}) over the whole trajectory.
log_reward	log R(x) at the trajectory's terminal state.
log_z	The current log Zθ estimate (LearnableScalar::value()).

Returns

The scalar loss, Δ(τ)².

grad_log_z()

float pulsatrix::TrajectoryBalanceLoss::grad_log_z

(

)

const

d(loss)/d(log Z) = 2Δ – feed directly into LearnableScalar::accumulate_grad.

grad_weight_for_log_pf()

float pulsatrix::TrajectoryBalanceLoss::grad_weight_for_log_pf

(

)

const

The weight w = -2Δ such that the per-trajectory-step policy-network gradient is w · (p[k] − 1{k==a_t}) – PolicyGradientLoss's exact softmax/log gradient identity shape, with this trajectory's Δ playing the role of the per-step return.

---

The documentation for this class was generated from the following file:

• include/pulsatrix/trajectory_balance_loss.hpp