Produced a series of reinforcement learning algorithms in Python to efficiently identify the optimal characteristics of prime membership advertisements through strategic sequential testing
Adapted these models to complex data settings, including sparse signals and heteroskedastic noise
Created realistic simulation structures to test best-advertisement identification algorithms (“best-arm identification” in the multi-armed bandit literature) against A/B testing and other competitors in high-dimensional data contexts
Applied Scientist, Sponsored Products
Amazon
Designed a reinforcement learning algorithm to better select advertisements sourced by a variety of machine learning algorithms using query features and partially observed customer behavior signals
Implemented the infrastructure for evaluating the advantages of modifications to advertisement sourcing algorithms on downstream outcomes in Spark
Ran off-line experiments testing the utility of my algorithm over a week of Amazon search query traffic and found that it was consistently capturing more than 3% of the high-quality ads missed by the current method