In machine learning, ensemble methods typically make better predictions and are less overfitted than individual models used in the ensemble.

Most of ensemble inference methods (with a notable exception of stacking, which is an example of ‣) don't create any now models and just do some deterministic computation on the outputs of the models in the ensemble, so we cannot think of them as "additional acts of creativity" which, as Deutsch wrote, are required to effectively combine several good explanations (models).

I suspect that ensemble methods work well despite being "dumb" for the same reason as Diversity in a workplace (or just about any system, actually: see ‣) is useful. Ensembling (and diversity) safeguard against biases and significantly reduces the loss where any particular explanation (opinion, model) is very wrong, while not hurting the performance much in typical cases.

Use an ensemble of models to prevent overfitting

Related:

See also: