Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 14-Day Trial for You or Your Team.

Learn More →

Rejoinder to ‘Multi‐armed bandit experiments in the online service economy’

Rejoinder to ‘Multi‐armed bandit experiments in the online service economy’ I thank Deepak Agarwal for his discussion and for his interesting example. It is true that massive scale poses a challenge to Thompson sampling because it can be difficult to simulate from the required posterior distribution. The model in Deepak's example is a mixed effects logistic regression, with random effects for ad‐level covariates and fixed effects for interactions between user and ad creative characteristics. This is a true ‘big data’ problem, which raises issues orthogonal to those I sought to highlight in the article, but which are nonetheless critical for ad serving at LinkedIn and most other large Internet companies. Deepak proposes to treat the fixed effects and the variance of the random effects as known and to do approximate Thompson sampling based on the conditional distribution of the random effects. Given today's computing constraints, this is a reasonable approach. Another possibility would be to parallelize the posterior simulations using a technique like consensus Monte Carlo . Consensus Monte Carlo partitions the data across many worker machines. A full MCMC is run on each worker machine to learn a local posterior distribution for A and σ 2 . The local posterior distributions are combined to find the global http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Applied Stochastic Models in Business and Industry Wiley

Rejoinder to ‘Multi‐armed bandit experiments in the online service economy’

Loading next page...
 
/lp/wiley/rejoinder-to-multi-armed-bandit-experiments-in-the-online-service-wlasQ0ksQ8

References (25)

Publisher
Wiley
Copyright
Copyright © 2015 John Wiley & Sons, Ltd.
ISSN
1524-1904
eISSN
1526-4025
DOI
10.1002/asmb.2107
Publisher site
See Article on Publisher Site

Abstract

I thank Deepak Agarwal for his discussion and for his interesting example. It is true that massive scale poses a challenge to Thompson sampling because it can be difficult to simulate from the required posterior distribution. The model in Deepak's example is a mixed effects logistic regression, with random effects for ad‐level covariates and fixed effects for interactions between user and ad creative characteristics. This is a true ‘big data’ problem, which raises issues orthogonal to those I sought to highlight in the article, but which are nonetheless critical for ad serving at LinkedIn and most other large Internet companies. Deepak proposes to treat the fixed effects and the variance of the random effects as known and to do approximate Thompson sampling based on the conditional distribution of the random effects. Given today's computing constraints, this is a reasonable approach. Another possibility would be to parallelize the posterior simulations using a technique like consensus Monte Carlo . Consensus Monte Carlo partitions the data across many worker machines. A full MCMC is run on each worker machine to learn a local posterior distribution for A and σ 2 . The local posterior distributions are combined to find the global

Journal

Applied Stochastic Models in Business and IndustryWiley

Published: Jan 1, 2015

There are no references for this article.