pulsatrix::GFlowNetTrajectory Struct Reference

One full sampled trajectory: every state the forward policy acted from, the action taken at each, and the trajectory-level quantities Trajectory Balance (and Detailed Balance/SubTB, in later missions) need. More...

#include <gflownet_trajectory.hpp>

Public Attributes
std::vector< Tensor >	states
	The state the policy acted from at each decision point, in order.
std::vector< int64_t >	actions
	The action index sampled at each corresponding entry of states.
float	sum_log_pf = 0.0f
	Σ_t log P_F(s_{t+1}|s_t) over the whole trajectory, including the final stop decision.
float	sum_log_pb = 0.0f
	Σ_t log P_B(s_t|s_{t+1}) over every real move (the stop action does not move the state, so it contributes no P_B term).
float	terminal_reward = 0.0f
	R(x) at the trajectory's terminal state.

Detailed Description

One full sampled trajectory: every state the forward policy acted from, the action taken at each, and the trajectory-level quantities Trajectory Balance (and Detailed Balance/SubTB, in later missions) need.

Note

states/actions deliberately are NOT enough on their own to recompute sum_log_pf/sum_log_pb – they exist so a training loop's second pass can re-forward() the policy network at each state (refreshing its single-most-recent-call cache) and apply the now-known per-step gradient. See mission_trajectory_balance_loss.md's Design section ("Two-pass rollout").

Member Data Documentation

actions

std::vector<int64_t> pulsatrix::GFlowNetTrajectory::actions

The action index sampled at each corresponding entry of states.

states

std::vector<Tensor> pulsatrix::GFlowNetTrajectory::states

The state the policy acted from at each decision point, in order.

sum_log_pb

float pulsatrix::GFlowNetTrajectory::sum_log_pb = 0.0f

Σ_t log P_B(s_t|s_{t+1}) over every real move (the stop action does not move the state, so it contributes no P_B term).

sum_log_pf

float pulsatrix::GFlowNetTrajectory::sum_log_pf = 0.0f

Σ_t log P_F(s_{t+1}|s_t) over the whole trajectory, including the final stop decision.

terminal_reward

float pulsatrix::GFlowNetTrajectory::terminal_reward = 0.0f

R(x) at the trajectory's terminal state.

---

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

• include/pulsatrix/gflownet_trajectory.hpp