Skip to content

Glossary

Terms used throughout this documentation. History matching borrows vocabulary from Bayesian calibration, design of experiments, and the emulation literature; the definitions below assume general familiarity with calibration but not with history matching itself.

ARD — Automatic Relevance Determination

A Gaussian Process kernel parameterization that fits a separate lengthscale to each input parameter. A short lengthscale means the output varies quickly with that parameter (it is "relevant"); a long one means the output barely depends on it. Used by the GPR emulator. (ARD reflects emulator sensitivity to the current wave's target features; the pairplot.png diagnostic ranks parameters differently — by overall constraint across all waves — so it does not use ARD.)

Bayes linear emulator

The default emulator (emulator_type='bayes_linear'). It combines a linear regression trend with a correlated-residual process (a squared-exponential kernel whose correlation lengths are fit by maximum likelihood) — effectively a Gaussian process with a linear mean. In the Bayes linear tradition it tracks the full predictive uncertainty (an adjusted expectation and variance) rather than just a point estimate, giving calibrated error bars at a fraction of the cost of GPR and with no TensorFlow dependency. Implemented in pure NumPy/SciPy.

\[ y(x) = \underbrace{\beta_0 + \textstyle\sum_j \beta_j x_j}_{\text{linear trend}} + r(x), \qquad r(\cdot) \sim \mathcal{GP}\!\left(0,\, \sigma^2 k\right), \qquad k(x, x') = \exp\!\left(-\sum_j \frac{(x_j - x'_j)^2}{\theta_j^2}\right) \]

with trend coefficients \(\beta\) (fit by OLS), residual variance \(\sigma^2\), and a per-parameter correlation length \(\theta_j\) fit by maximum likelihood.

Emulator (surrogate)

A fast statistical approximation of the (expensive) simulator, trained on a modest number of simulation runs. The emulator predicts simulator output — with an uncertainty estimate — at new parameter values for a tiny fraction of the cost. History matching uses emulator predictions, not the simulator itself, to rule out parameter space.

Fano factor

The variance-to-mean ratio of an output across the sampled points. The automatic feature-selection method {'method': 'fano'} uses it to pick the most informative outputs to emulate: outputs whose value changes a lot relative to their mean carry more signal for constraining parameters. See also mean_sq_z and feature/output.

Feature / output

A scalar summary of a simulation run that you have a corresponding observation for (e.g. peak incidence, attack rate, cases in week 1). The two words are used interchangeably: "feature selection" chooses which model outputs to emulate in a given wave. Each selected feature gets its own emulator.

GLM — Generalized Linear Model

An emulator (emulator_type='glm') that fits a linear predictor through a link function \(g\):

\[ g\!\left(\mathbb{E}[y(x)]\right) = \beta_0 + \sum_j \beta_j x_j \]

With the default Gaussian family (\(g\) = identity) it reduces to the linear emulator; with the Poisson family (\(g = \log\), link='poisson') it models non-negative count outputs.

GPR — Gaussian Process Regression

A flexible, nonlinear emulator (emulator_type='gpr') built on GPflow/TensorFlow with ARD kernels. Best for nonlinear response surfaces and excellent uncertainty estimates, at a higher computational cost than the linear emulators.

\[ y(x) \sim \mathcal{GP}(m,\, k), \qquad k(x, x') = \sigma_f^2 \exp\!\left(-\tfrac{1}{2}\sum_j \frac{(x_j - x'_j)^2}{\ell_j^2}\right) \]

with a constant mean \(m\), signal variance \(\sigma_f^2\), observation-noise variance \(\sigma_n^2\), and a separate lengthscale \(\ell_j\) per parameter (ARD).

Implausibility

A score measuring how inconsistent a parameter value is with one observation, expressed in standard deviations. For output \(f\) at parameter \(x\):

\[ I(x) = \frac{\left|\, \mathbb{E}[f(x)] - z \,\right|}{\sqrt{\operatorname{Var}_{\text{emulator}} + \operatorname{Var}_{\text{obs}} + \operatorname{Var}_{\text{disc}}}} \]

where \(z\) is the observed target and the denominator combines the emulator's predictive variance, the observation's variance, and an optional model discrepancy variance accounting for structural mismatch between the simulator and reality. (The discrepancy is supplied as a standard deviation \(\sigma_{\text{disc}}\) and enters as its square, \(\operatorname{Var}_{\text{disc}} = \sigma_{\text{disc}}^2\)model_discrepancy, default 0.) With several outputs, the implausibility of a point is the maximum over all outputs, so a point must be consistent with every constraint to survive. A point is ruled out when its implausibility exceeds the threshold.

Implausibility threshold

The implausibility cutoff above which a parameter value is discarded (implausibility_threshold, default 3.0). The default of 3 follows the Vysochanskij–Petunin inequality (popularized as Pukelsheim's "three-sigma rule"): for any unimodal distribution with finite variance — not just a normal one — at least ~95% of the probability mass lies within 3 standard deviations of the mean. So an implausibility above 3 is strong, distribution-free evidence that the point is inconsistent with the observation.

LHS — Latin Hypercube Sampling

A space-filling experimental design that spreads sample points evenly across every parameter's range. The default sampling strategy ('lhs'); better coverage than uniform random sampling for the same number of points.

Linear emulator

The simplest emulator (emulator_type='linear'): an ordinary least-squares fit

\[ y(x) = \beta_0 + \sum_j \beta_j x_j + \varepsilon, \qquad \varepsilon \sim \mathcal{N}(0, \sigma^2) \]

with predictive uncertainty taken from the OLS standard errors. Fast and interpretable, but assumes the response is linear in the parameters — see GLM for non-Gaussian outputs, or Bayes linear and GPR for nonlinear response surfaces.

Maximin

A space-filling criterion that maximizes the minimum distance between sample points, pushing them apart for even coverage. Used as an LHS criterion ({'type': 'lhs', 'criterion': 'maximin'}) and as the final "thinning" stage of the ray_resample pipeline, where it selects a well-spread subset from a larger candidate pool.

mean_sq_z

The default automatic feature-selection method ({'method': 'mean_sq_z'}): mean squared z-score, i.e. how far each output sits from its target in standard-deviation units, averaged across samples. Outputs that are far from their target carry the most information for ruling out parameter space. See also Fano factor.

NROY — Not Ruled Out Yet

The region of parameter space that history matching has not yet shown to be implausible — the central output of the method. Instead of a single best-fit point, history matching returns this set of all parameter values that could plausibly have produced the observed data. The NROY region shrinks wave by wave. NROY samples are points drawn from it (via get_nroy_samples()), and the NROY fraction is the share of fresh prior samples that fall inside it — a convergence diagnostic that falls toward zero as the calibration tightens.

Prior

The initial parameter space — the bounds you start from, before any wave has ruled anything out. History matching progressively carves the NROY region out of the prior.

Ray sampling / ray_resample

A method for drawing NROY samples efficiently when the NROY region is a tiny fraction of the prior and plain rejection sampling becomes slow. The ray_resample pipeline (inspired by the hmer R package) is four stages: LHS → ray sampling (drawing points along lines connecting known NROY points) → importance sampling → maximin thinning. Faster than pure LHS at low acceptance rates, but can over-represent region boundaries; see the NROY sampling methods tutorial.

Trajectory selection

A post-calibration step for stochastic models: having found which parameters are plausible, select specific (parameter set, random seed) pairs whose simulated trajectories match the observed data, using importance resampling weighted by a pseudo-likelihood. See the trajectory selection tutorial.

Wave

One iteration of the history matching loop — sample, simulate, select features, train emulators, filter. "Wave" and "iteration" are used interchangeably; the on-disk output is organized into wave1/, wave2/, etc. Each wave starts from the NROY region left by the previous one, so the searched space contracts with every wave.