Back to search

Boston University

Temporal aspects of adaptive online learning: continuity and representation

Abstract

dc:description.abstract

Adaptive online learning, in a very broad sense, is the study of sequential decision making beyond the worst case. Compared to their classical minimax counterparts, adaptive algorithms typically require less manual tuning, while provably performing better in benign environments, or with prior knowledge. This dissertation presents new techniques for designing these algorithms. The central theme is the emphasis on the temporal nature of the problem, which has not received enough attention in the literature. The first part of the dissertation focuses on temporal continuity. While modern online learning almost exclusively studies a discrete time repeated game, it is shown that designing algorithms can be simplified, and in certain cases optimized, by scaling the game towards a continuous time limit and solving the obtained differential equation. Concretely, we develop comparator adaptive algorithms for Online Convex Optimization, achieving optimal static regret bounds in the vanilla setting and its variant with switching costs. The benefits are extended to another classical online learning problem called Learning with Expert Advice. The second part of the dissertation focuses on temporal representation. Different from the first part, here we consider the general objective of dynamic regret minimization, which forms the foundation of time series forecasting. It is shown that by introducing temporal features, the task can be transformed to static regret minimization on a user-specified representation space with growing dimension. Drawing novel connections to wavelet features, we develop a simple algorithm improving the state-of-the-art dynamic regret bound achieved by more sophisticated approaches. An application is the online fine-tuning of a black-box time series forecaster.

Degree

thesis:*
Name thesis:degree_name
Doctor of Philosophy
Level thesis:degree_level
doctoral
Discipline thesis:degree_discipline
Systems Engineering
Grantor
Boston University
Year dc:date.issued
2023

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Zhang, Zhiyu
Advisors dc:contributor.advisor
  • Paschalidis, Ioannis Ch.
  • Cutkosky, Ashok
ORCID dc:identifier.orcid

Subjects

dc:subject × 1

Rights

Language dc:language.iso
en_US

Identifiers

dc:identifier.*
Handle dc:identifier.uri
https://hdl.handle.net/2144/46657

Chain of custody

source
Harvested from
Boston University
Base URL
open.bu.edu/oai/request
Last updated
2026-11-18
Source record
OAI-PMH GetRecord
citation

Zhang, Zhiyu. Temporal aspects of adaptive online learning: continuity and representation. doctoral thesis, Boston University, 2023. https://hdl.handle.net/2144/46657