GradientKernel<T>
Models & Types · AiDotNet.Kernels
Kernel that incorporates gradient observations for GPs with derivative information.
Implements: IKernelFunction<T>
For Beginners
Sometimes you have access not just to function values f(x), but also to gradient information ∇f(x) (the rate of change in each direction).
This is common in:
- Bayesian optimization: When evaluating an expensive function, you might also
get gradient information “for free” (e.g., via automatic differentiation)
- Physics-informed ML: Physical laws often constrain derivatives
- Adjoint methods: Gradients are computed alongside function values
The GradientKernel extends a base kernel to model both values and gradients. If f(x) is a GP with kernel k(x, x’), then:
- Cov(f(x), f(x’)) = k(x, x’)
- Cov(∂f/∂xᵢ, f(x’)) = ∂k(x, x’)/∂xᵢ
- Cov(∂f/∂xᵢ, ∂f/∂x’ⱼ) = ∂²k(x, x’)/∂xᵢ∂x’ⱼ
This lets you use gradient observations to improve predictions, especially when gradient observations are cheaper than function observations.
How It Works
Usage: Create with a base kernel (RBF, Matern, etc.) that supports gradients. The resulting kernel operates on extended vectors: [x; gradient_dim_flag; is_gradient_obs].
Constructors
| Constructor | Summary |
|---|---|
GradientKernel(int inputDim, GradientKernelType<T> kernelType, double lengthScale, double outputScale) |
Initializes a new GradientKernel. |
Properties
| Property | Summary |
|---|---|
GradientKernelType<T> KernelType { get; } |
Gets the kernel type. |
IKernelFunction<T> BaseKernel { get; } |
Gets the base kernel. |
double LengthScale { get; } |
Gets the length scale. |
int InputDim { get; } |
Gets the input dimensionality. |
Methods
| Method | Summary |
|---|---|
T Calculate(Vector<T> x1, Vector<T> x2) |
Calculates the kernel value between two points (value-value, value-gradient, or gradient-gradient). |
Vector<T> CreateGradientObservation(Vector<T> input, int gradientDim) |
Creates an extended vector for a gradient observation. |
Vector<T> CreateValueObservation(Vector<T> input) |
Creates an extended vector for a value observation. |