Appraise#

Appraisal stage of the Neighbourhood Algorithm as in Sambridge 1999(II). Based on an approximation of the objective surface created by the Direct Search stage of the NA, the Appraisal stage resamples the approximation without needing to re-evaluate the objective function. This is a fast and efficient way to obtain more samples from the objective surface.

class neighpy.appraise.NAAppraiser(n_resample: int, n_walkers: int = 1, searcher: neighpy.search.NASearcher | None = None, initial_ensemble: numpy.ndarray[Any, numpy.dtype[numpy._typing._generic_alias.ScalarType]] | None = None, log_ppd: numpy.ndarray[Any, numpy.dtype[numpy._typing._generic_alias.ScalarType]] | None = None, bounds: Optional[Tuple[Tuple[float, float], ...]] = None, verbose: bool = True, seed: int | None = None)#

Can be initialized with either a NASearcher object or the initial_ensemble, log_ppd and bounds. If initialising with a NASearcher object, the NASearcher.objectives attribute will be assumed to be the negative log posterior probability density. For example, if the objective function used in the NASearcher is something like

\[\Phi(\mathbf{m}) = (\mathbf{d} - \mathbf{G}(\mathbf{m}))^TC_d^{-1}(\mathbf{d} - \mathbf{G}(\mathbf{m}))\]

i.e. the negative log-likelihood of a Gaussian data misfit. This will be passed to NAAppraiser as log_ppd = -NASearcher.objectives.

Parameters:

  • n_resample (int) – The number of resamples to use for the appraisal.

  • n_walkers (int) – The number of walkers to use in parallel.

  • searcher (NASearcher) – The searcher object used to generate the initial ensemble and log_ppd. If this is not None, the initial_ensemble, log_ppd and bounds will be taken from this object.

  • initial_ensemble (NDArray) – The initial ensemble of samples. Ignored if searcher is not None.

  • log_ppd (NDArray) – The log posterior probability values for each sample. Ignored if searcher is not None.

  • bounds (Tuple[Tuple[float, float], ...]) – The bounds of the parameter space. Ignored if searcher is not None.

  • verbose (bool) – Whether to display a progress bar.

  • seed (int) – Seed for the random number generator.

run(save: bool = True, start_fraction: float = 0.5) → None#

Perform the appraisal stage of the Neighbourhood Algorithm. Calculates a few basic MC integrals (mean, covariance and their errors).

Parameters:

  • save – bool - whether to save the new samples from the appraisal. Set to :code:False to save memory, or if you only want a mean and covariance.

  • start_fraction – float - the fraction of the best cells to start the random walks from. This is used to avoid walking in low probability regions. Only used if n_walkers > 1.

Populates the following attributes:

  • samples (NDArray) - the samples generated by the appraisal (if save=True).

  • mean (NDArray) - the mean of the samples.

  • sample_mean_error (NDArray) - the error in the mean.

  • covariance (NDArray) - the covariance of the samples.

  • sample_covariance_error (NDArray) - the error in the covariance.