Struct PerformancePredictor 

pub struct PerformancePredictor<T: Float + Debug + Send + Sync + 'static> { /* private fields */ }

Performance prediction system.

Wraps the real learned predictor in crate::evaluation::PerformancePredictor (a ridge-regularised linear model over a deterministic architecture feature vector, trained online from observed evaluations). Earlier releases returned a constant 0.6 here.

Implementations

impl<T: Float + Debug + Default + Send + Sync + 'static> PerformancePredictor<T>

pub fn new(config: &EvaluationConfig<T>) -> Result<Self>

pub fn predict( &mut self, architecture: &OptimizerArchitecture<T>, ) -> Result<EvaluationResults<T>>

Predict the performance of an architecture without running it.

Delegates to the learned model, which extracts a deterministic feature vector from the architecture and scores it; the returned confidence interval widens when little training data has been seen.

pub fn update_training_data( &mut self, results: &[SearchResult<T>], ) -> Result<()>

Feed observed evaluations back into the model.

Each result performs one online gradient step on the predictor’s weights, and the running prediction accuracy (1 - mean absolute error over the retained history) is refreshed.

pub fn prediction_accuracy(&self) -> T

Current estimated prediction accuracy in [0, 1].

pub fn confidence_threshold(&self) -> T

Minimum accuracy at which the predictor should be trusted in place of a full evaluation.

pub fn model_type(&self) -> &PredictorType

Kind of model backing this predictor.

pub fn training_sample_count(&self) -> usize

Number of (architecture, evaluation) pairs observed so far.

Trait Implementations

impl<T: Debug + Float + Debug + Send + Sync + 'static> Debug for PerformancePredictor<T>

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.