Chatty Stochastic Multi-Armed Bandits

Abstract

This thesis uses a variant of the classic stochastic multi-armed bandit framework to improve the user experience in an online chat application by selecting conversation starters. While the traditional algorithm would converge on the `optimal' conversation starter and use it for every conversation, this novel version of the algorithm attempts to provide new conversation starters for each user while still attempting to maximize the conversation quality. This thesis examines the empirical behavior of such an algorithm in a web application deployed at Princeton University.