saealib.GenCtrlSwitcher#

class saealib.GenCtrlSwitcher(*, gm_max=5, gm_min=0, update_rate=0.5, metric='spearman', initial_error=0.5)[source]#

Bases: AccuracyBasedSurrogateSwitcher[int]

Adjust gen_ctrl via exponential smoothing (Repický et al., 2017).

The surrogate error is estimated from an accuracy metric, exponentially smoothed, then linearly mapped to gen_ctrl:

\[ \begin{align}\begin{aligned}\varepsilon_t = (1-r_u)\varepsilon_{t-1} + r_u\varepsilon_{\text{new}}\\g_m = \operatorname{round}((1-\varepsilon_t) \cdot g_{m,\max})\end{aligned}\end{align} \]

For Spearman rho and R2, epsilon = 0.5 * (1 - value) maps the metric to [0, 1]. For RMSE the value is used directly (assumed to be in [0, 1]). When accuracy is unavailable the smoothed state is unchanged and the current estimate is returned.

Parameters:

  • gm_max (int) – Upper bound for gen_ctrl. Default: 5.

  • gm_min (int) – Lower bound for gen_ctrl. Default: 0.

  • update_rate (float) – Smoothing rate r_u in (0, 1]. Default: 0.5.

  • metric (str) – Metric key. Default: "spearman".

  • initial_error (float) – Initial smoothed error before the first observation. Default: 0.5 (neutral).

smoothed_error#

Current exponentially smoothed error estimate. Inspect to monitor the internal state during an iterate() run.

Type:

float

Methods

__init__

switch

Update smoothed error and return new gen_ctrl.

Method Details

GenCtrlSwitcher.__init__(*, gm_max=5, gm_min=0, update_rate=0.5, metric='spearman', initial_error=0.5)[source]#
Parameters:

  • gm_max (int)

  • gm_min (int)

  • update_rate (float)

  • metric (str)

  • initial_error (float)

Return type:

None

GenCtrlSwitcher.switch(accuracy)[source]#

Update smoothed error and return new gen_ctrl.

Parameters:

accuracy (SurrogateAccuracy | None)

Return type:

int