Package smile.regression

Class GaussianProcessRegression<T>

java.lang.Object
smile.regression.GaussianProcessRegression<T>
Type Parameters:
T - the data type of model input objects.
All Implemented Interfaces:
Serializable, ToDoubleFunction<T>, Regression<T>
public class GaussianProcessRegression<T> extends Object implements Regression<T>
Gaussian Process for Regression. A Gaussian process is a stochastic process whose realizations consist of random values associated with every point in a range of times (or of space) such that each such random variable has a normal distribution. Moreover, every finite collection of those random variables has a multivariate normal distribution.

A Gaussian process can be used as a prior probability distribution over functions in Bayesian inference. Given any set of N points in the desired domain of your functions, take a multivariate Gaussian whose covariance matrix parameter is the Gram matrix of N points with some desired kernel, and sample from that Gaussian. Inference of continuous values with a Gaussian process prior is known as Gaussian process regression.

The fitting is performed in the reproducing kernel Hilbert space with the "kernel trick". The loss function is squared-error. This also arises as the kriging estimate of a Gaussian random field in spatial statistics.

A significant problem with Gaussian process prediction is that it typically scales as O(n3). For large problems (e.g. n > 10,000) both storing the Gram matrix and solving the associated linear systems are prohibitive on modern workstations. An extensive range of proposals have been suggested to deal with this problem. A popular approach is the reduced-rank Approximations of the Gram Matrix, known as Nystrom approximation. Subset of Regressors (SR) is another popular approach that uses an active set of training samples of size m selected from the training set of size n > m. We assume that it is impossible to search for the optimal subset of size m due to combinatorics. The samples in the active set could be selected randomly, but in general we might expect better performance if the samples are selected greedily w.r.t. some criterion. Recently, researchers had proposed relaxing the constraint that the inducing variables must be a subset of training/test cases, turning the discrete selection problem into one of continuous optimization.

Experimental evidence suggests that for large m the SR and Nystrom methods have similar performance, but for small m the Nystrom method can be quite poor. Also, embarrassments can occur like the approximated predictive variance being negative. For these reasons we do not recommend the Nystrom method over the SR method.

References

  1. Carl Edward Rasmussen and Chris Williams. Gaussian Processes for Machine Learning, 2006.
  2. Joaquin Quinonero-candela, Carl Edward Ramussen, Christopher K. I. Williams. Approximation Methods for Gaussian Process Regression. 2007.
  3. T. Poggio and F. Girosi. Networks for approximation and learning. Proc. IEEE 78(9):1484-1487, 1990.
  4. Kai Zhang and James T. Kwok. Clustered Nystrom Method for Large Scale Manifold Learning and Dimension Reduction. IEEE Transactions on Neural Networks, 2010.
See Also:

  • Field Details

    • kernel

      public final MercerKernel<T> kernel
      The covariance/kernel function.

    • regressors

      public final T[] regressors
      The regressors.

    • w

      public final double[] w
      The linear weights.

    • mean

      public final double mean
      The mean of responsible variable.

    • sd

      public final double sd
      The standard deviation of responsible variable.

    • noise

      public final double noise
      The variance of noise.

    • L

      public final double L
      The log marginal likelihood, which may be not available (NaN) when the model is fit with approximate methods.

  • Constructor Details

    • GaussianProcessRegression

      public GaussianProcessRegression(MercerKernel<T> kernel, T[] regressors, double[] weight, double noise)
      Constructor.
      Parameters:
      kernel - Kernel function.
      regressors - The regressors.
      weight - The weights of regressors.
      noise - The variance of noise.

    • GaussianProcessRegression

      public GaussianProcessRegression(MercerKernel<T> kernel, T[] regressors, double[] weight, double noise, double mean, double sd)
      Constructor.
      Parameters:
      kernel - Kernel function.
      regressors - The regressors.
      weight - The weights of regressors.
      noise - The variance of noise.
      mean - The mean of responsible variable.
      sd - The standard deviation of responsible variable.

    • GaussianProcessRegression

      public GaussianProcessRegression(MercerKernel<T> kernel, T[] regressors, double[] weight, double noise, double mean, double sd, Matrix.Cholesky cholesky, double L)
      Constructor.
      Parameters:
      kernel - Kernel function.
      regressors - The regressors.
      weight - The weights of regressors.
      noise - The variance of noise.
      mean - The mean of responsible variable.
      sd - The standard deviation of responsible variable.
      cholesky - The Cholesky decomposition of kernel matrix.
      L - The log marginal likelihood.

  • Method Details

    • predict

      public double predict(T x)
      Description copied from interface: Regression
      Predicts the dependent variable of an instance.
      Specified by:
      predict in interface Regression<T>
      Parameters:
      x - an instance.
      Returns:
      the predicted value of dependent variable.

    • predict

      public double predict(T x, double[] estimation)
      Predicts the mean and standard deviation of an instance.
      Parameters:
      x - an instance.
      estimation - an output array of the estimated mean and standard deviation.
      Returns:
      the estimated mean value.

    • query

      public GaussianProcessRegression<T>.JointPrediction query(T[] samples)
      Evaluates the Gaussian Process at some query points.
      Parameters:
      samples - query points.
      Returns:
      The mean, standard deviation and covariances of GP at query points.

    • toString

      public String toString()
      Overrides:
      toString in class Object

    • fit

      public static GaussianProcessRegression<double[]> fit(double[][] x, double[] y, Properties params)
      Fits a regular Gaussian process model.
      Parameters:
      x - the training dataset.
      y - the response variable.
      params - the hyper-parameters.
      Returns:
      the model.

    • fit

      public static <T> GaussianProcessRegression<T> fit(T[] x, double[] y, MercerKernel<T> kernel, Properties params)
      Fits a regular Gaussian process model.
      Type Parameters:
      T - the data type of samples.
      Parameters:
      x - the training dataset.
      y - the response variable.
      kernel - the Mercer kernel.
      params - the hyper-parameters.
      Returns:
      the model.

    • fit

      public static <T> GaussianProcessRegression<T> fit(T[] x, double[] y, MercerKernel<T> kernel, double noise)
      Fits a regular Gaussian process model by the method of subset of regressors.
      Type Parameters:
      T - the data type of samples.
      Parameters:
      x - the training dataset.
      y - the response variable.
      kernel - the Mercer kernel.
      noise - the noise variance, which also works as a regularization parameter.
      Returns:
      the model.

    • fit

      public static <T> GaussianProcessRegression<T> fit(T[] x, double[] y, MercerKernel<T> kernel, double noise, boolean normalize, double tol, int maxIter)
      Fits a regular Gaussian process model.
      Type Parameters:
      T - the data type of samples.
      Parameters:
      x - the training dataset.
      y - the response variable.
      kernel - the Mercer kernel.
      noise - the noise variance, which also works as a regularization parameter.
      normalize - the flag if normalize the response variable.
      tol - the stopping tolerance for HPO.
      maxIter - the maximum number of iterations for HPO. No HPO if maxIter <= 0.
      Returns:
      the model.

    • fit

      public static <T> GaussianProcessRegression<T> fit(T[] x, double[] y, T[] t, MercerKernel<T> kernel, Properties params)
      Fits an approximate Gaussian process model by the method of subset of regressors.
      Type Parameters:
      T - the data type of samples.
      Parameters:
      x - the training dataset.
      y - the response variable.
      t - the inducing input, which are pre-selected or inducing samples acting as active set of regressors. In simple case, these can be chosen randomly from the training set or as the centers of k-means clustering.
      kernel - the Mercer kernel.
      params - the hyper-parameters.
      Returns:
      the model.

    • fit

      public static <T> GaussianProcessRegression<T> fit(T[] x, double[] y, T[] t, MercerKernel<T> kernel, double noise)
      Fits an approximate Gaussian process model by the method of subset of regressors.
      Type Parameters:
      T - the data type of samples.
      Parameters:
      x - the training dataset.
      y - the response variable.
      t - the inducing input, which are pre-selected or inducing samples acting as active set of regressors. In simple case, these can be chosen randomly from the training set or as the centers of k-means clustering.
      kernel - the Mercer kernel.
      noise - the noise variance, which also works as a regularization parameter.
      Returns:
      the model.

    • fit

      public static <T> GaussianProcessRegression<T> fit(T[] x, double[] y, T[] t, MercerKernel<T> kernel, double noise, boolean normalize)
      Fits an approximate Gaussian process model by the method of subset of regressors.
      Type Parameters:
      T - the data type of samples.
      Parameters:
      x - the training dataset.
      y - the response variable.
      t - the inducing input, which are pre-selected or inducing samples acting as active set of regressors. In simple case, these can be chosen randomly from the training set or as the centers of k-means clustering.
      kernel - the Mercer kernel.
      noise - the noise variance, which also works as a regularization parameter.
      normalize - the option to normalize the response variable.
      Returns:
      the model.

    • nystrom

      public static <T> GaussianProcessRegression<T> nystrom(T[] x, double[] y, T[] t, MercerKernel<T> kernel, Properties params)
      Fits an approximate Gaussian process model with Nystrom approximation of kernel matrix.
      Type Parameters:
      T - the data type of samples.
      Parameters:
      x - the training dataset.
      y - the response variable.
      t - the inducing input, which are pre-selected for Nystrom approximation.
      kernel - the Mercer kernel.
      params - the hyper-parameters.
      Returns:
      the model.

    • nystrom

      public static <T> GaussianProcessRegression<T> nystrom(T[] x, double[] y, T[] t, MercerKernel<T> kernel, double noise)
      Fits an approximate Gaussian process model with Nystrom approximation of kernel matrix.
      Type Parameters:
      T - the data type of samples.
      Parameters:
      x - the training dataset.
      y - the response variable.
      t - the inducing input, which are pre-selected for Nystrom approximation.
      kernel - the Mercer kernel.
      noise - the noise variance, which also works as a regularization parameter.
      Returns:
      the model.

    • nystrom

      public static <T> GaussianProcessRegression<T> nystrom(T[] x, double[] y, T[] t, MercerKernel<T> kernel, double noise, boolean normalize)
      Fits an approximate Gaussian process model with Nystrom approximation of kernel matrix.
      Type Parameters:
      T - the data type of samples.
      Parameters:
      x - the training dataset.
      y - the response variable.
      t - the inducing input, which are pre-selected for Nystrom approximation.
      kernel - the Mercer kernel.
      noise - the noise variance, which also works as a regularization parameter.
      normalize - the option to normalize the response variable.
      Returns:
      the model.