spires_opt_score

Struct controlling how the optimizer scores candidate reservoir configurations.

Definition

enum spires_metric_kind { SPIRES_METRIC_AUROC = 0, SPIRES_METRIC_AUPRC = 1 };
 
struct spires_opt_score {
    double lambda_var;
    double lambda_cost;
    int    metric;
    int    _reserved_i0;
};

Fields

Composite Score Formula

The optimizer ranks candidates by:

score = metric_mean - lambda_var * metric_std - lambda_cost * relative_cost

where:

• metric_mean is the average of the selected metric across all seeds in a budget stage.
• metric_std is the standard deviation across seeds.
• relative_cost is a normalized measure of reservoir size (proportional to num_neurons).

Example

/* Optimize AUROC with moderate variance penalty, low cost penalty */
struct spires_opt_score score = {
    .lambda_var  = 0.1,
    .lambda_cost = 0.01,
    .metric      = SPIRES_METRIC_AUROC
};

/* Optimize AUPRC for imbalanced datasets, no cost penalty */
struct spires_opt_score score = {
    .lambda_var  = 0.2,
    .lambda_cost = 0.0,
    .metric      = SPIRES_METRIC_AUPRC
};

Notes

• AUROC vs. AUPRC. AUROC is appropriate when positive and negative classes are roughly balanced. For highly imbalanced datasets (e.g., rare-event detection), AUPRC provides a more informative measure because it is sensitive to precision at low recall.
• Setting lambda_var to 0. Disables the variance penalty. The optimizer will select the configuration with the highest mean metric, regardless of how much performance varies across random seeds.
• Setting lambda_cost to 0. Disables the cost penalty. The optimizer will not penalize larger reservoirs, which may lead to selecting configurations with many neurons.
• Reserved fields. Always zero-initialize _reserved_i0.

See Also

• spires_optimize — consumes this struct.
• spires_opt_budget — budget stage configuration.
• spires_opt_result — result struct containing the final scores.
• Scoring Metrics — practical guidance on metric selection.