Low-Regret Learning

05 Jun 2016 22:01

Yet Another Inadequate Placeholder.

A.k.a. "prediction with expert advice", "individual-sequence prediction", "on-line learning". (That last name bugs me. Strictly speaking, one could have an on-line learning algorithm which didn't aim at low regret but at, say, low risk, and indeed the old uses of stochastic optimization for regression aimed implicitly at just that.)

See also: Ensemble Methods in Machine Learning; Forecasting Non-Stationary Processes; Learning Theory; Universal Prediction