Perturbation techniques in online learning and optimization

In this chapter we give a new perspective on so-called perturbation methods that have been applied in a number of different fields, but in particular for adversarial online learning problems. We show that the classical algorithm known as Follow The Perturbed Leader (FTPL) can be viewed through the lens of stochastic smoothing, a tool that has proven popular within convex optimization. We prove bounds on regret for several online learning settings, and provide generic tools for analyzing perturbation algorithms. We also consider the so-called bandit setting, where the feedback to the learner is significantly constrained, and we show that near-optimal bounds can be achieved as long as a simple condition on the perturbation distribution is met.