Online Program Home
  My Program

Abstract Details

Activity Number: 425 - SPEED: Reliable Statistical Learning and Data Science
Type: Contributed
Date/Time: Tuesday, August 1, 2017 : 3:05 PM to 3:50 PM
Sponsor: Section on Statistical Learning and Data Science
Abstract #325327
Title: Using Bandit Algorithms on Changing Reward Rates
Author(s): Jeffrey Roach*
Companies: OpenMail
Keywords: multi-armed bandit ; online advertising ; Thompson Sampling ; changing reward rates
Abstract:

Optimizing a dynamic online advertising system can be difficult for a multi-armed bandit. We utilized a sliding window of data along with a time decay to control overconfidence in the performance of a feature. We combined these techniques with a Thompson Sampling bandit to balance our explore/exploit strategy, minimize shocks during transition periods, and handle no clear winner situations efficiently resulting in ~5% increase in overall payout.


Authors who are presenting talks have a * after their name.

Back to the full JSM 2017 program

 
 
Copyright © American Statistical Association