Kernels & Means

Simple Kernel & Mean

You may already seen, we can specify a kernel function like this(same for mean fucntions):

k = pyGPs.cov.RBF( log_ell=-1., log_sigma=0. )

There are several points need to be noticed:

  1. Most parameters are initilized in their logorithms. This is because we need to make sure they are positive during optimization. e.g. Here length scale and signal variance should always be positive.
  2. Most kernel functions have a scalar in front, namely signal variance(set by log_sigma)
  3. If you will do optimization later anyway, you can just leave parameters to be default

Some Special Cases

  1. For some kernels/means, number of hyperparameters depends on the dimension of input data. You can either enter the dimension, which use default values:

    m = pyGPs.mean.Linear( D=x.shape[1] )
    

    or you can initialze with the exact hyperparameters, you should enter as a list, one element for each dimension

    m = pyGPs.mean.Linear( alpha_list=[0.2, 0.4, 0.3] )
    
    All these “hyp-dim-dependent” functions are:
    • pyGPs.mean.Linear
    • pyGPs.cov.RBFard
    • pyGPs.cov.LINard
    • pyGPs.cov.RQard
  2. For pyGPs.cov.RBFunit(), its signal variance is always 1 (because of unit magnitude). Therefore this function do not have a hyperparameter of “signal variance”.

  3. pyGPs.cov.Poly() has three parameters, where hyperparameters are:
    • c -> inhomogeneous offset
    • sigma -> signal deviation
    however,
    • d -> order of polynomial will be treated as normal parameter, i.e. will not be trained
  4. Explicitly set pyGPs.cov.Noise is not necessary, because noise are already added in likelihood.

Composite Kernels & Meams

Adding and muliplying Kernels(Means) is really simple:

k = pyGPs.cov.Linear() * pyGPs.cov.RBF()
k = 0.5 * pyGPs.cov.Linear() + pyGPs.cov.RBF()

Scalar will also be treated as a hyperparameter. For example, k = s1 * k1 + s2 * k2, then the list of hyperparameters is hyp = [s1, k1.hyp, s2, k2.hyp]. Scalar is passed in logorithm domain such that it will always be positive during optimization.

Beside + / * , there is also a power operator for mean functions:

m = ( pyGPs.mean.One() + pyGPs.mean.Linear(alpha_list=[0.2]) )**2

Precomputed Kernel Matrix

In certain cases, you may have a precomputed kernel matrix, but its non-trivial to write down the exact formula of kernel functions. Then you can specify your kernel in the following way. A precomputed kernel also fits with other kernels. In other words, it can also be composited as the way other kernels functions do.

k = pyGPs.cov.Pre(M1, M2)

M1 and M2 are your precomputed kernel matrix,

where,

M1 is a matrix with shape number of training points plus 1 by number of test points
  • cross covariances matrix (train by test)
  • last row is self covariances (diagonal of test by test)
M2 is a square matrix with number of training points for each dimension
  • training set covariance matrix (train by train)

A precomputed kernel can also be composited with other kernels. You need to explictly add scalar for pyGPs.cov.Pre().

k = 0.5*pyGPs.cov.Pre(M1, M2) + pyGPs.cov.RBF()