Theory and applications of competitive prediction

摘要: Predicting the future is an important purpose of machine learning research. In online learning, predictions are given sequentially rather than all at once. People wish to make sensible decisions in many situations everyday life, whether month-by-month, day-by-day, or minute-by-minute. competitive prediction, made by a set experts and learner. The quality measured loss function. goal learner reliable under any circumstances. compares his with best from ensures that performance not much worse. this thesis general methodology described provide algorithms strong guarantees for prediction problems. Specific attention paid square function, widely used assess predictions. Four types sets considered thesis: finite number free (which required follow strategy), following strategies finite-dimensional spaces, infinite-dimensional Hilbert Banach spaces. power illustrated derivations various algorithms. Two core approaches explored Aggregating Algorithm Defensive Forecasting. These close each other interesting cases. However, Forecasting more covers some problems which cannot be solved using Algorithm. specific often computationally efficient. empirical properties new validated on artificial real world data sets. areas where can applied emphasized.