Skip to contents

Query the most informative unlabeled candidates

Source: R/active_query.R
al_query.Rd

Given a fitted Active-Learning model, picks the next n candidates to sample according to a trade-off strategy. All strategies use a greedy batch-mode selection: the second and later picks within the same call see the previously picked points as if they were already labeled, so the batch remains internally diverse.

Usage

al_query(
  model,
  candidates,
  n = 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
)

Arguments

model

A edaphos_al_model.

candidates

Data frame of unlabeled candidate locations. Must contain model$covariates. Rows with NA in any covariate are ignored.

n

Integer, batch size.

strategy

One of "hybrid", "uncertainty", "diverse", "cost".

alpha

Numeric in [0, 1], weight of uncertainty vs diversity in hybrid / cost.

quantiles

Length-2 numeric with the lower/upper probabilities used to measure the QRF interval width. Defaults to c(0.1, 0.9).

base

Numeric length-2 vector c(x, y); required by strategy = "cost".

cost_weight

Numeric, weight of the cost penalty in "cost".

physics_gate

Optional function function(candidates, predicted_mean) -> logical that returns TRUE for physically feasible candidates and FALSE for infeasible ones. Infeasible rows are excluded from the greedy selection, linking Pillar 5 to Pillar 2 — see al_physics_gate_piml() for a ready-made gate driven by a PIML profile fit.

Value

Integer vector of row indices in candidates that form the next batch, in selection order.

Strategies

  • "uncertainty" — highest QRF prediction-interval width (pure exploitation of model uncertainty).

  • "diverse" — max-min distance in the standardised covariate space from both the current labeled set and the already-picked batch members (pure exploration).

  • "hybrid" — convex combination alpha * uncertainty + (1 - alpha) * diversity on 0-1 scaled scores. The recommended default.

  • "cost" — "hybrid" minus cost_weight * cost, where cost is the 0-1 normalised Euclidean distance to a logistical base (x, y). Use this to steer an autonomous sampler (drone, rover) towards points it can physically reach with limited energy.