Skip to contents

Search Instance

Source: R/SearchInstance.R
SearchInstance.Rd

Container for a search problem that extends bbotk's bbotk::EvalInstance to support both optimization and active learning. Holds the objective, search space, archive, and terminator, and provides the evaluation loop mechanics.

Unlike bbotk::OptimInstance, this class:

  • Supports codomains with "learn" tags (in addition to minimize/maximize)

  • Does not compute objective_multiplicator

  • Does not assign a "result" (best point)

Creates a new SearchInstance.

Evaluates a batch of points and adds results to the archive.

Resets the instance for a fresh search.

Printer.

Arguments

objective

(bbotk::Objective)
The objective to evaluate. Can be any bbotk Objective subclass including our ObjectiveDataset and ObjectiveLearner.

search_space

(paradox::ParamSet)
Optional restricted search space. If NULL:

  • If the domain contains no TuneTokens, uses the whole domain

  • If the domain contains TuneTokens, derives search space from them Cannot be supplied if the domain already contains TuneTokens.

terminator

(bbotk::Terminator)
When to stop the search. Uses bbotk terminators.

archive

(bbotk::ArchiveBatch)
Optional pre-existing archive. If NULL, creates a new one.

check_values

(logical(1))
Whether to validate points against search_space before evaluation.

callbacks

(list() of mlr3misc::Callback)
Optional callbacks to hook into the evaluation loop.

xdt

(data.table)
Points to evaluate, one row per configuration.

...

(ignored).

Value

Invisibly returns the ydt (codomain values).

Details

The search instance serves as the "problem specification" for custom search loops. It owns the archive and handles evaluation mechanics.

If you want to run a full MBO-style loop, prefer using bbotk's bbotk::OptimInstanceBatchSingleCrit / bbotk::OptimInstanceBatchMultiCrit together with mlr3mbo::OptimizerMbo. For codomains containing "learn" targets, use ResultAssignerNull to disable assigning a "best" result.

Since bbotk's Codomain now natively supports "learn" tags, codomains are passed directly to ArchiveBatch without conversion. Note that calling archive$best() or archive$nds_selection() will error if the codomain contains only "learn" targets, which is the correct behavior.

Callbacks

Callbacks can be registered to hook into the evaluation loop. The following stages are supported:

  • on_optimizer_before_eval: Called before evaluating a batch

  • on_optimizer_after_eval: Called after evaluating a batch

Termination

Before each batch evaluation, the terminator is checked. If terminated, a search_terminated_error condition is raised. This can be caught with tryCatch(..., search_terminated_error = function(e) ...).