data_models.likelihoods.api
data_models.likelihoods.api
Classes
| Name | Description |
|---|---|
| Likelihood | Model of how observations scatter around the underlying response. |
| GaussianLikelihood | Observations are the response plus Gaussian noise with mean zero and one |
| SparseOutlierGaussianLikelihood | Gaussian noise plus an extra noise variance for observations treated as |
Likelihood
data_models.likelihoods.api.Likelihood()Model of how observations scatter around the underlying response.
It decides how much of the spread in the data is attributed to measurement noise rather than to the response itself.
GaussianLikelihood
data_models.likelihoods.api.GaussianLikelihood()Observations are the response plus Gaussian noise with mean zero and one variance shared by all points.
The noise variance is fitted along with the kernel hyperparameters; noise_prior regularises it and noise_constraint bounds it. By default there is no prior and the noise is at least 1e-4.
The noise is that of the outputs as the model sees them, after the surrogate’s output scaler (standardized by default), so priors and bounds refer to that scale and not to the raw measurements.
SparseOutlierGaussianLikelihood
data_models.likelihoods.api.SparseOutlierGaussianLikelihood()Gaussian noise plus an extra noise variance for observations treated as outliers, so that they do not distort the fit.
Models with different numbers of outliers (0%, 5%, 10%, 20%, … of the observations) are fitted, and one is chosen by Bayesian model comparison, under a prior that favours few outliers (relevance pursuit, Ament et al. 2024, https://arxiv.org/abs/2410.24222). noise_prior and noise_constraint refer to the noise shared by all observations. Works with one output only.
What counts as an outlier depends on how flexible the model is: a kernel with short lengthscales can explain almost any point. Bound the lengthscales from below, e.g. with ROBUSTGP_LENGTHSCALE_CONSTRAINT().