Skip to contents

Diffusion-posterior-driven AL (Pilar 9 x Pilar 5)

Source: R/bridges_active_learning_v3.R
al_query_diffusion.Rd

Ranks candidate cells by the standard deviation of DDPM posterior samples. High-SD cells are the most uncertain map locations; labelling there delivers the largest reduction in model entropy per sample.

Usage

al_query_diffusion(
  dm_fit,
  conditioning = NULL,
  n_samples = 16L,
  candidate_cells = NULL,
  n_select = 10L,
  combine = c("sd", "sd_x_mean_abs"),
  seed = NULL
)

Arguments

dm_fit

An edaphos_dm_fit from dm_fit().

conditioning

Optional (n_samples, cond_dim) matrix (one row per posterior draw); when NULL the model is sampled unconditionally.

n_samples

Integer; number of posterior draws.

candidate_cells

Optional integer matrix with columns row, col restricting the candidate pool to accessible sites. Defaults to the full H x W grid.

n_select

Integer; how many cells to return.

combine

One of "sd" or "sd_x_mean_abs" to weight SD by the absolute expected value (prioritises cells where the model thinks the SOC is BOTH high and uncertain).

seed

RNG seed.

Value

A data frame of class edaphos_al_diffusion_query sorted by descending score.