Struct SelectionNetwork 

pub struct SelectionNetwork<A: Float> { /* private fields */ }

Neural network for optimizer selection

A single-hidden-layer perceptron with ReLU activation and a softmax output, trained with cross-entropy loss. SelectionNetwork::train backpropagates through both layers, so the hidden representation is learned rather than frozen at its initialization.

Implementations

impl<A: Float + ScalarOperand + Debug + FromPrimitive + Send + Sync> SelectionNetwork<A>

pub fn new(input_size: usize, hidden_size: usize, num_optimizers: usize) -> Self

Create a new selection network

pub fn from_parameters( input_weights: Array2<A>, input_bias: Array1<A>, output_weights: Array2<A>, output_bias: Array1<A>, ) -> Result<Self>

Build a network from explicit parameters

Useful for reproducible experiments, checkpoint restore, and gradient verification, where the pseudo-random initialization of SelectionNetwork::new is not acceptable.

Errors

Returns an error if the four parameter arrays do not describe a consistent input -> hidden -> output topology.

pub fn hidden_size(&self) -> usize

Size of the hidden layer

pub fn input_size(&self) -> usize

Number of input features the network expects

pub fn num_outputs(&self) -> usize

Number of optimizer classes the network scores

pub fn input_weights(&self) -> &Array2<A>

Read-only view of the hidden-layer (input -> hidden) weights

pub fn output_weights(&self) -> &Array2<A>

Read-only view of the output-layer (hidden -> logits) weights

pub fn input_bias(&self) -> &Array1<A>

Read-only view of the hidden-layer biases

pub fn output_bias(&self) -> &Array1<A>

Read-only view of the output-layer biases

pub fn forward(&self, features: &Array1<A>) -> Result<Array1<A>>

Forward pass to get optimizer probabilities

pub fn average_loss( &self, features: &[Array1<A>], optimizer_labels: &[usize], ) -> Result<A>

Average cross-entropy loss over a labelled dataset

Useful for monitoring that SelectionNetwork::train is actually reducing the objective. Returns zero for an empty dataset.

pub fn train( &mut self, features: &[Array1<A>], optimizer_labels: &[usize], learning_rate: A, epochs: usize, ) -> Result<()>

Train the network on historical data with full backpropagation

Runs plain SGD on the cross-entropy loss for epochs passes over the data. Every parameter is updated: the output layer from the softmax delta p − onehot(label), and the hidden layer from that delta propagated back through W₂ᵀ and gated by the ReLU derivative (1 where the pre-activation is positive, 0 elsewhere).

Arguments

• features - Input feature vectors
• optimizer_labels - Index of the correct optimizer for each feature vector
• learning_rate - SGD step size
• epochs - Number of passes over the dataset

Errors

Returns an error if the feature and label counts disagree, if a label is out of range, or if a feature vector has the wrong length.

Trait Implementations

impl<A: Debug + Float> Debug for SelectionNetwork<A>

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

Formats the value using the given formatter.