GaussianProcess

The Boulder Opal Toolkits are currently in beta phase of development. Breaking changes may be introduced.

class GaussianProcess(length_scale_bounds=None, seed=None)

The Gaussian process optimizer.

Parameters
  • length_scale_bounds (np.ndarray, optional) – The per-parameter length scale bounds on the test points. The bounds must be a NumPy array of shape (parameter count, 2) where the trailing axis are the bounds for each parameter (with the lower bound first, followed by the upper bound).

  • seed (int, optional) – Seed for the random number generator. Use this option to generate deterministic results from the optimizer.

Notes

The Gaussian process is defined by the kernel

\[k({\mathbf x}_j, {\mathbf x}_k) = \exp \left(-\frac{1}{2} ( {\mathbf x}_j - {\mathbf x}_k )^\top \Sigma^{-2} ( {\mathbf x}_j - {\mathbf x}_k )\right) ,\]

where \({\mathbf x}_j\) is an \(n\)-dimensional vector representing the \(j\)-th test point, \(\Sigma= {\rm diag}(l_1, \cdots, l_n)\) is an \(n \times n\) diagonal matrix, and \(\{ l_j \}\) are the length scales. The length scales are tuned while training the model, within the bounds set by the length_scale_bounds parameter. Roughly speaking, the amount a parameter needs to change to impact the optimization cost should lie within the length scale bounds.

It’s recommended to provide non-zero cost_uncertainty to optimize when using this optimizer, otherwise you might encounter a numerical error when the optimizer tries to fit the kernel with your input data. If the error persists, try increasing the cost_uncertainty value or decreasing the minimum length scale bound. However, such numerical error is also an indication that your data might not be suitable to be modelled by a Gaussian process, and in that case, consider using a different closed-loop optimizer.

For more detail on Gaussian processes see Gaussian process on Wikipedia.

Methods

create_optimizer

Return the initialized optimizer.