Struct PerformancePredictor 

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

Performance predictor

Implements a genuinely-learned, feature-based online ridge-regression predictor. An OptimizerArchitecture is mapped to a fixed-length numeric feature vector and scored by a linear model squashed through a logistic function into [0, 1]. The model is updated online from observed evaluation results.

Removed scaffolding

Earlier revisions carried a large tree of neural-network / feature-pipeline / prediction-cache types around this model: ModelArchitecture, RegularizationParameters, LearningRateSchedule, EarlyStoppingState, FeatureExtractor, FeatureEngineeringPipeline, FeatureScaling, FeatureInteractions, PolynomialFeatures, FeatureSelection, FeatureCache, TrainingMetadata, ResourceUsageRecord, DataSplits, PredictionCache, PredictionResult, CacheStatistics, CacheConfig, UncertaintyEstimator, UncertaintyParameters, CalibrationData and CalibrationCurve. Every one of them was constructed once, stored, and never read again — not one influenced a single prediction, and none could be built or inspected from outside the module. They are gone rather than left as a public API that does nothing; what remains is exactly the state the model uses, and all of it is readable through the accessors below.

Implementations

impl<T: Float + Debug + Default + Send + Sync> PerformancePredictor<T>

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

pub fn model_type(&self) -> &PredictorModelType

Model family this predictor implements.

pub fn epochs_trained(&self) -> usize

Number of online updates applied so far.

pub fn training_loss_history(&self) -> &[T]

Squared error of every online update, in application order.

pub fn observation_count(&self) -> usize

Number of (features -> target) pairs the model has been trained on.

pub fn training_rmse(&self) -> Option<T>

Root-mean-square error of the current weights over every training pair recorded so far, or None before any pair has been observed.

This is a resubstitution error, not a held-out estimate: it says how well the model now fits what it has already seen. It is what makes the stored training features and targets observable rather than write-only.

pub fn confidence_level(&self) -> T

Confidence level used to widen the reported prediction interval.

pub fn set_confidence_level(&mut self, level: T) -> Result<()>

Set the confidence level used for the reported prediction interval.

Rejects values outside the open interval (0, 1): a level of 0 or 1 is not a confidence level, and silently clamping it would make the reported interval mean something other than what the caller asked for.

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

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

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.