test.kernel.grad: Test an Analytic Gradient Against a Numeric Gradient

In mattdneal/gaussianProcess: Gaussian Process Fitting and Prediction

Description

Creates a numerical approximation to the gradient function of a kernel (using create.numeric.grad) and compares it to an analytic gradient function. If the maximum relative difference exceeds 0.01, returns the failing inputs, otherwise returns TRUE.

Usage

test.kernel.grad(k, k.grad, hyper.param.names, additional.params,
  repetitions = 1000, dimensions = NULL)

Arguments

k	a kernel function
k.grad	the analytic gradient to test
hyper.param.names	names of the kernel's hyperparameters
additional.params	any additional parameters of the kernel
repetitions	number of repetitions to attempt

Value

If the gradients match, returns TRUE. If a mismatch is found, returns a list with named entries:

• a, b, hyper.params - the parameters passed to the gradient functions

• grad - return value of the analytic gradient

• grad.num - return value of the numerical gradient

• max.diff - the maximum relative difference observed between the two gradient functions so far (including the failed set of inputs)

• i - the number of inputs tested (including the failed set of inputs)

mattdneal/gaussianProcess documentation built on May 21, 2019, 12:58 p.m.