Closed-loop Autonomous Active Learning for soil mapping

Runs the full Pillar 5 closed loop:

1. fit a Quantile Regression Forest on the current labeled set;

2. query the next batch of candidates via al_query();

3. label them through a user-supplied oracle (or a simulation oracle, when the reference values are already in candidates);

4. append + refit;

5. repeat until the sampling budget is exhausted or the OOB RMSE falls below stop_rmse.

Usage

al_loop(
  labeled,
  candidates,
  target,
  covariates,
  coords = NULL,
  budget = 30L,
  batch = 5L,
  strategy = c("hybrid", "uncertainty", "diverse", "cost"),
  alpha = 0.7,
  quantiles = c(0.1, 0.9),
  base = NULL,
  cost_weight = 0.3,
  physics_gate = NULL,
  oracle = NULL,
  stop_rmse = NULL,
  num.trees = 500L,
  verbose = TRUE
)

Arguments

labeled

Data frame with the initial labeled sample.

candidates

Data frame of unlabeled candidate locations. When oracle is NULL, candidates[[target]] must already contain the true values (simulation mode).

target

Character, name of the target column.

covariates

Character vector of covariate column names.

coords

Optional length-2 character vector naming x/y columns.

budget

Integer, total number of samples the loop is allowed to query (in addition to the initial set).

batch

Integer, number of samples to query per iteration.

strategy

Query strategy — see al_query().

alpha, quantiles, base, cost_weight, physics_gate

Forwarded to al_query(). See al_physics_gate_piml() for a PIML-backed physics gate that couples this Pillar 5 loop to a Pillar 2 fit.

oracle

Optional labelling function function(samples) -> numeric. When NULL, a simulation oracle reads the true target from candidates.

stop_rmse

Optional early-stop threshold on OOB RMSE.

num.trees

Integer, forest size used at every refit.

verbose

Logical; print per-iteration diagnostics.

Value

A edaphos_al_model whose $history contains one entry per iteration (iter 0 is the initial fit).

Details

The oracle argument is the extension point that lets the same code drive simulation, lab-pipeline integration, or on-board Edge-AI decision making on a drone/rover (the Pillar-5 endgame).

Examples

# \donttest{
  if (requireNamespace("sp", quietly = TRUE)) {
    data(meuse, package = "sp")
    d <- stats::na.omit(meuse[, c("x", "y", "dist", "elev",
                                  "ffreq", "soil", "lead")])
    d$ffreq <- as.numeric(d$ffreq)
    d$soil  <- as.numeric(d$soil)
    set.seed(1)
    seed_idx  <- al_initial_design(d, c("dist", "elev", "ffreq", "soil"),
                                   n = 15, iter = 500)
    m <- al_loop(
      labeled    = d[seed_idx, ],
      candidates = d[-seed_idx, ],
      target     = "lead",
      covariates = c("dist", "elev", "ffreq", "soil"),
      coords     = c("x", "y"),
      budget     = 20, batch = 5,
      strategy   = "hybrid", verbose = FALSE
    )
    al_history(m)
  }
#>   iter n_labeled rmse_oob mean_uncertainty
#> 1    0        15 53.49527               NA
#> 2    1        20 63.21191            242.6
#> 3    2        25 63.28423            227.0
#> 4    3        30 62.15733            234.4
#> 5    4        35 67.83500            198.2
# }