GPWithMCMC<T>

Models & Types · AiDotNet.GaussianProcesses

Gaussian Process with Markov Chain Monte Carlo inference for full Bayesian treatment.

Inheritance: ModelBase<T, Matrix<T>, Vector<T>> → GaussianProcessBase<T> → GPWithMCMC<T>

Implements: IGeneratedParameterRegistrar<T>

For Beginners

Standard GP makes point estimates of hyperparameters (kernel lengthscale, output variance, noise variance). MCMC provides a fully Bayesian treatment by sampling from the posterior distribution of hyperparameters, giving better uncertainty quantification.

Instead of finding a single “best” lengthscale, MCMC explores many plausible lengthscales and averages predictions across all of them. This is more robust when:

  • You have limited data
  • The hyperparameters are uncertain
  • You need accurate uncertainty estimates

The implementation uses Slice Sampling, which automatically adapts to the target distribution without requiring careful tuning of step sizes.

Constructors

Constructor Summary
GPWithMCMC(IKernelFunction<T> kernel, int numSamples, int burnIn, int thinning, double logLengthscalePriorMean, double logLengthscalePriorStd, double logVariancePriorMean, double logVariancePriorStd, int? seed) Initializes a GP with MCMC inference.

Properties

Property Summary
IKernelFunction<T> Kernel { get; } Gets the kernel function.
bool IsTrained { get; } Gets whether the GP is trained.
int NumStoredSamples { get; } Gets the number of stored MCMC samples.

Methods

Method Summary
Dictionary<string, ValueTuple<double, double>> GetPosteriorStatistics() Computes posterior statistics for hyperparameters.
List<double[]> GetSamples() Gets the MCMC samples [lengthscale, outputVariance, noiseVariance].
ValueTuple<T, T> Predict(Vector<T> x) Predicts at a single test point with full posterior averaging.
ValueTuple<Vector<T>, Vector<T>> PredictBatch(Matrix<T> Xtest) Predicts at multiple test points.
void Fit(Matrix<T> X, Vector<T> y) Fits the GP to training data using MCMC sampling.
void UpdateKernel(IKernelFunction<T> newKernel) Updates the kernel (not supported for MCMC GP).