Where does ensemble collide?

A statistical ensemble is a large imagined collection of copies of the same physical system, each copy standing for one microscopic arrangement, or microstate, that the system could occupy under fixed conditions. Gibbs introduced this device in his 1902 Elementary Principles in Statistical Mechanics, where averaging a quantity across the whole collection, rather than following one system through time, recovers the thermodynamic properties of matter. The three standard versions, the microcanonical, canonical, and grand canonical ensembles, differ in what they hold fixed, such as energy or particle number. Here the referent is a conceptual population of system copies used to compute averages, not a set of weather runs, a group of trained models, or a group of musicians.[1]

Unit

One imagined copy of the system in a definite microstate; the ensemble is the whole weighted collection of such copies

Recovery means

Not applicable

Measured by

Formed as a probability distribution over microstates, from which averages and fluctuations of observable quantities are computed

An ensemble in weather prediction is a set of separate forecast runs started from slightly different initial conditions, or built with different model settings, so that the spread among them estimates how uncertain the forecast is. Epstein's 1969 Stochastic dynamic prediction first treated a forecast as the evolution of a probability distribution rather than a single trajectory, and Leith's 1974 Theoretical skill of Monte Carlo forecasts showed that running a modest collection of perturbed forecasts approximates that distribution in practice. Each member is a full run of the same equations, and the collection, not any one run, is the product. Here the referent is a family of competing weather trajectories, not a population of microstates, a set of combined statistical models, or a group of performers.[2][3]

Unit

One full model run, or member; the ensemble is the set of members and their spread

Recovery means

Not applicable

Measured by

Summarised by the ensemble mean and the spread across members, and calibrated against observed outcomes

In machine learning, an ensemble is a set of separately trained models whose individual predictions are combined, usually by voting or averaging, into a single prediction that is typically more accurate than any one member. Hansen and Salamon's 1990 Neural network ensembles introduced the term for this setting, showing that a collection of networks voting together made fewer errors than a single network. Breiman's 1996 Bagging predictors and later work such as Dietterich's 2000 Ensemble methods in machine learning established standard ways of building and explaining such combinations. Here the referent is a group of predictive models pooled into one, not a set of physical system copies, a spread of weather forecasts, or a group of musicians.[4][5][6]

Unit

A collection of trained base models plus the rule that combines them

Recovery means

Not applicable

Measured by

Evaluated by the combined model's predictive accuracy relative to its individual members