UC Berkeley

Online learning and convex optimization algorithms have become essential tools for solving problems in modern machine learning, statistics and engineering. Many algorithms for online learning and convex optimization can be interpreted as a discretization of a continuous time process, and studying the continuous time dynamics offers many advantages: the analysis is often simpler and more elegant in continuous time, it provides insights and leads to new interpretations of the discrete process, and streamlines the design of new algorithms, obtained by deriving the dynamics in continuous time, then discretizing. In this thesis, we apply this paradigm to two problems: the study of decision dynamics for online learning in games, and the design and analysis of accelerated methods for convex optimization.

In the first part of the thesis, we study online learning dynamics for a class of games called non-atomic convex potential games, which are used for example to model congestion in transportation and communication networks. We make a connection between the discrete Hedge algorithm for online learning, and an ODE on the simplex, known as the replicator dynamics. We study the asymptotic properties of the ODE, then by discretizing the ODE and using results from stochastic approximation theory, we derive a new class of online learning algorithms with asymptotic convergence guarantees. We further give a more refined analysis of these dynamics and their convergence rates. Then, using the Hedge algorithm as a model of decision dynamics, we pose and study two related problems: the problem of estimating the learning rates of the Hedge algorithm given observations on its sequence of decisions, and the problem of optimal control under Hedge dynamics.

In the second part, we study first-order accelerated dynamics for constrained convex optimization. We develop a method to design an ODE for the problem using an inverse Lyapunov argument: we start from an energy function that encodes the constraints of the problem and the desired convergence rate, then design an ODE tailored to that energy function. Then, by carefully discretizing the ODE, we obtain a family of accelerated algorithms with optimal rate of convergence. This results in a unified framework to derive and analyze most known first-order methods, from gradient descent and mirror descent to their accelerated versions. We give different interpretations of the ODE, inspired from physics and statistics. In particular, we give an averaging interpretation of accelerated dynamics, and derive simple sufficient conditions on the averaging scheme to guarantee a given rate of convergence. We also develop an adaptive averaging heuristic that empirically speeds up the convergence, and in many cases performs significantly better than popular heuristics such as restarting.