ComputeGAE()'s two outputs: the advantage estimate per step and the critic's regression target per step. More...

#include <gae.hpp>

Collaboration diagram for pulsatrix::GAEResult:

Public Attributes
Tensor	advantages
	`A_t` – the GAE(gamma, lambda) advantage estimate, the actor's per-step weight.

Tensor	returns
	`A_t + V(s_t)` – the critic's regression target. Deliberately not the raw Monte-Carlo return RolloutBuffer::compute_returns() produces; see ComputeGAE().

Detailed Description

ComputeGAE()'s two outputs: the advantage estimate per step and the critic's regression target per step.

Note: Plain data, no behavior – the same small named return-struct precedent as ReparamGrad (reparameterize.hpp), StepResult (environment.hpp) and RolloutBatch (rollout_buffer.hpp). A std::pair<Tensor, Tensor> would make every call site read .first/.second for two tensors that are emphatically not interchangeable.; Both are (N, 1), row t of each belonging to stored step t – the same leading dimension as the rewards/dones/values passed in.

Member Data Documentation

Tensor pulsatrix::GAEResult::advantages

A_t – the GAE(gamma, lambda) advantage estimate, the actor's per-step weight.

Tensor pulsatrix::GAEResult::returns

A_t + V(s_t) – the critic's regression target. Deliberately not the raw Monte-Carlo return RolloutBuffer::compute_returns() produces; see ComputeGAE().

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