CS 229 T / STATS 231 : Statistical Learning Theory
This lecture will show the rates of FTRL on the concrete examples the authors discussed earlier to build some intuition, and introduce the entropic regularizer, a useful regularizer when they have to make actions that are probability distributions.
Abstract
Last time, we discussed how the Follow the Leader algorithm does not perform very well as an online convex optimization algorithm, and how we could introduce the Follow the Regularized Leader algorithm to fix some of the shortcomings of FTL. In this lecture we will show the rates of FTRL on the concrete examples we discussed earlier to build some intuition. To use FTRL on the expert problem we will introduce the entropic regularizer, a useful regularizer when we have to make actions that are probability distributions. Recall that on each timestep FTRL chooses