Predictive learning with online algorithms

Abstract

that contains online learning models and randomized online algorithms is designed for time-based peak shaving problems. The thesis defines this kind of online problem as a time-based demand-uncertainty online problem, including ski rental, peak shaving, cloud resource allocation, and so on. This thesis also explores the usage of extra modules that work with the original framework to improve the performance of online algorithms in time-based demand-uncertainty online problems. Mathematical proof in this thesis proves that these modules can obtain a better bounded competitive ratio, consistency, and robustness. We also conducted experiments on simulated data sets and real-world data sets to prove the effectiveness of these modules.