ARDKernel<T>

Models & Types · AiDotNet.Kernels

Implements the Automatic Relevance Determination (ARD) kernel with per-dimension length scales.

Implements: IKernelFunction<T>

For Beginners

The ARD (Automatic Relevance Determination) kernel is a powerful extension of the standard RBF kernel that uses a different length scale for each input dimension.

In mathematical terms: k(x, x’) = σ² × exp(-0.5 × Σᵢ (xᵢ - x’ᵢ)² / lᵢ²)

Where:

  • σ² is the overall variance (signal variance)
  • lᵢ is the length scale for dimension i

How It Works

Why is ARD important?

  1. Feature Selection: The length scales tell you which features are important:
  • Small lᵢ → Feature i is very relevant (small changes matter a lot)
  • Large lᵢ → Feature i is less relevant (can be ignored)
  • Very large lᵢ → Feature i is essentially irrelevant
  1. Dimensionality Handling: In high-dimensional problems, not all features matter equally.

ARD automatically figures out which ones are important.

  1. Interpretability: After training, inspect the length scales to understand

which features drive your predictions.

  1. Regularization: By learning length scales, the model avoids overfitting

to irrelevant features.

How it works:

  • Start with some initial length scales
  • Optimize them by maximizing the log marginal likelihood
  • Features with large optimized length scales can be pruned

Constructors

Constructor Summary
ARDKernel(Vector<T> lengthScales) Initializes a new ARD kernel with specified length scales.
ARDKernel(Vector<T> lengthScales, double variance) Initializes a new ARD kernel with per-dimension length scales and a signal variance.
ARDKernel(int numDimensions, double lengthScale, double variance) Initializes a new ARD kernel with uniform length scales.

Methods

Method Summary
T Calculate(Vector<T> x1, Vector<T> x2) Calculates the ARD kernel value between two vectors.
Vector<T> GetLengthScales() Gets the current length scales.
double GetVariance() Gets the signal variance.