Use this URL to cite or link to this record in EThOS: https://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.779388
Title: Optimization methods for structured machine learning problems
Author: Tsipinakis, Nikolaos
ISNI:       0000 0004 7965 0849
Awarding Body: UCL (University College London)
Current Institution: University College London (University of London)
Date of Award: 2019
Availability of Full Text:
Access from EThOS:
Full text unavailable from EThOS. Please try the link below.
Access from Institution:
Abstract:
Solving large-scale optimization problems lies at the core of modern machine learning applications. Unfortunately, obtaining a sufficiently accurate solution quickly is a difficult task. However, the problems we consider in many machine learning applications exhibit a particular structure. In this thesis we study optimization methods and improve their convergence behavior by taking advantage of such structures. In particular, this thesis constitutes of two parts: In the first part of the thesis, we consider the Temporal Difference learning (TD) problem in off-line Reinforcement Learning (RL). In off-line RL, it is typically the case that the number of samples is small compared to the number of features. Therefore, recent advances have focused on efficient algorithms to incorporate feature selection via `1-regularization which effectively avoids over-fitting. Unfortunately, the TD optimization problem reduces to a fixed-point problem where convexity of the objective function cannot be assumed. Further, it remains unclear whether existing algorithms have the ability to offer good approximations for the task of policy evaluation and improvement (either they are non-convergent or do not solve the fixed-point problem). In this part of the thesis, we attempt to solve the `1- regularized fixed-point problem with the help of Alternating Direction Method of Multipliers (ADMM) and we argue that the proposed method is well suited to the structure of the aforementioned fixed-point problem. In the second part of the thesis, we study multilevel methods for large-scale optimization and extend their theoretical analysis to self-concordant functions. In particular, we address the following issues that arise in the analysis of second-order optimization methods based either on sampling, randomization or sketching: (a) the analysis of the iterates is not scale-invariant and (b) lack of global fast convergence rates without restrictive assumptions. We argue that, with the analysis undertaken in this part of the thesis, the analysis of randomized second-order methods can be considered on-par with the analysis of the classical Newton method. Further, we demonstrate how our proposed method can exploit typical spectral structures of the Hessian that arise in machine learning applications to further improve the convergence rates.
Supervisor: Not available Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID: uk.bl.ethos.779388  DOI: Not available
Share: