Online convex optimization: algorithms, learning, and duality

• Autor(a) principal: Portella, Victor Sanches
• Data de Publicação: 2019
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Biblioteca Digital de Teses e Dissertações da USP
• Texto Completo: http://www.teses.usp.br/teses/disponiveis/45/45134/tde-05062019-165924/
• topic: Algorithms; Algoritmos; Convex; Convex Optimization; Convexa; Dualidade; Duality; Learning; Online; Online convex optimization; Otimização; Otimização convexa; Otimização convexa online

Online Convex Optimization (OCO) is a field in the intersection of game theory, optimization, and machine learning which has been receiving increasing attention due to its recent applications to a wide range of topics such as complexity theory and graph sparsification. Besides the usually simple description and implementation of OCO algorithms, a lot of this recent success is due to a deepening of our understanding of the OCO setting and their algorithms by using cornerstone ideas from convex analysis and optimization such as the powerful results from convex duality theory. In this text we present a mostly self-contained introduction to the field of online convex optimization. We first describe the online learning and online convex optimization settings, proposing an alternative way to formalize both of them so we can make formal claims in a clear and unambiguous fashion while not cluttering the readers understanding. We then present an overview of the main concepts of convex analysis we use, with a focus on building intuition. With respect to algorithms for OCO, we first present and analyze the Adaptive Follow the Regularized Leader (AdaFTRL) together with an analysis which relies mainly on the duality between strongly convex and strongly smooth functions. We then describe the Adaptive Online Mirror Descent (AdaOMD) and the Adaptive Dual Averaging (AdaDA) algorithms and analyze both by writing them as special cases of the AdaFTRL algorithm. Additionally, we show simple sufficient conditions for Eager and Lazy Online Mirror Descent (the non-adaptive counter-parts of AdaOMD and AdaDA) to be equivalent. We also present the well-known AdaGrad and Online Newton Step algorithms as special cases of the AdaReg algorithm, proposed by Gupta, Koren, and Singer, which is itself a special case of the AdaOMD algorithm. We conclude by taking a bird\'s-eyes view of the connections shown throughout the text, forming a ``genealogy\'\' of OCO algorithms, and discuss some possible path for future research.