Probabilistic reduction

Probabilistic reducers drop pending parameter combinations using completed model metrics. They formalize the research step of inspecting past results and deciding which values to stop trying. They can reduce completed work; whether that improves the experiment depends on the metric, sample size and search space.

Argument	Input	Description
`reduction_method`	str or callable	Reducer name or callable
`reduction_interval`	int	Number of permutations after which reduction is applied
`reduction_window`	int	the look-back window for reduction process
`reduction_threshold`	float	The threshold at which reduction is applied
`reduction_metric`	str	The metric to be used for reduction
`minimize_loss`	bool	`reduction_metric` is a loss

Prerequisites and procedure

Install the callback’s backend and execute the Scan minimal example. Choose a recorded numeric metric and its direction before configuring reduction. A writable run directory retains result and audit artifacts.

Select a reducer and a metric present in the callback history.
Set the interval and look-back window according to the completed observations needed for a useful decision.
Run the bounded API example below.
Inspect completed rows and the remaining queue before scaling to a research experiment.

The reduction arguments are passed to Scan().

Examples below use the held-out Iris setup in Scan → Minimal Example. Run that setup first; it defines scan_object, p, input_model, x, y, x_val, y_val, x_test, and y_test.

reduced = talos.Scan(x, y, p, input_model, 'reduced',
          x_val=x_val, y_val=y_val, seed=17, disable_progress_bar=True,
          reduction_method='correlation',
          reduction_interval=2,
          reduction_window=2,
          reduction_threshold=0.2,
          reduction_metric='val_loss',
          minimize_loss=True)

The example configures a Spearman decision every two completed trials, with a two-trial look-back and threshold 0.2, to minimize validation loss. The correlation implementation requires at least three finite target observations and variation in the target, so this deliberately small window does not prune. Increase the window and search size for a statistical experiment.

For each parameter value observed in the window, Talos builds a presence indicator and correlates it with the selected metric. The strongest unfavorable finite correlation meeting reduction_threshold selects a value to remove from pending combinations. When minimizing loss, a positive association with higher loss is unfavorable; when maximizing a metric, a negative association is unfavorable. Completed trials remain recorded.

Available reducers

Choose one of the below in reduction_method:

correlation (same as spearman)
spearman
kendall
pearson
trees
forrest
local_strategy: dynamically changing local strategy; see local strategy

For meaningful statistical reduction, use a larger sweep/window than this bounded API example. The native SFD core also provides Sanity, Saturation, Correlation, Focus and Budget reducers; see SFD and CLI.

Expected result and failure boundaries

reduced.data contains completed metrics and candidate values; reduced.param_object.param_index represents pending legacy rows. A pruning decision reduces the pending set and enters the run audit. Too few observations, a constant target or no varying candidate indicators produce no useful statistical reduction.

A metric absent from the result table raises an error. Choose minimize_loss to match the metric’s meaning: the name alone does not decide its direction. Tree reducers use a different decision mechanism; reduction_threshold governs the correlation family rather than every reducer. The accepted forest reducer name retains the historical spelling forrest.

Prerequisites and procedure​

Available reducers​

Expected result and failure boundaries​

Read next​

Prerequisites and procedure

Available reducers

Expected result and failure boundaries

Read next