CMSV-TOCS: Scott Clark 2013-04-09

Optimally Learning for Fun and Profit

I will introduce and explain the multi-armed bandit problem within an intuitive and mathematical framework. I will motivate why it is important to Yelp and other web companies and explain how it can be used for more effective and efficient A/B/N testing and site improvement. I will discuss several strategies and compare their relative merits. Next I will show results comparing several methods in situations likely to be encountered in the real world. I will talk about extensions to standard algorithms that allow for adaptation to specific problems and scenarios seen at Yelp and other web-based companies and varying ways to attack these problems. I will conclude with some real world examples of the algorithms and techniques discussed, including some current and ongoing research.

I am going to be showing some cool machine learning tricks centered around the multi-armed bandit problem. I will talk about how they work and show some results. While there will be some math on the slides, I will motivate everything intuitively with many, many graphs and pictures. It should be accessable to anyone with any math background (or lack thereof).

Speaker Bio:

Scott recently finished his PhD in Applied Mathematics at Cornell University and is now a Software Engineer on the Ad Targeting team at Yelp. He is enjoying the transition from academia to industry and is trying to learn and experience as much as possible in the Bay Area tech community.

He enjoys coming up with and implementing new algorithms to solve difficult problems more efficiently with emphasis on parallelization and exploiting huge data sets using statistics and machine learning.