Deep Reinforcement Learning for Query Optimization in Database Systems

Explore the potential of machine learning in database management systems through this colloquium talk by Ryan Marcus from MIT. Delve into two innovative approaches to query optimization using deep reinforcement learning: Neo and Bao. Discover how Neo combines tree convolution neural networks with value iteration techniques to replace traditional query optimizers, achieving significant performance improvements. Learn about Bao's strategy for dynamic workloads, which trains an agent to guide existing query optimizers using a contextual multi-armed bandit framework. Gain insights into the broader implications of applying machine learning to systems problems, including the potential benefits and challenges of fully learned systems. Understand the speaker's background in researching learned systems, focusing on query optimization, data storage, and indexing, as well as his work at Intel Labs and academic journey.

Allen School Colloquium: Ryan Marcus (MIT)