Mathematical Foundations of Modern Learning Problems
Abstract
This proposal aims to bring together techniques from Machine Learning and Control. We propose to employ the toolkit of statistical learning for the problems of control and identification from finite trajectories of observations. In particular, our research program is to show that sequential complexities, developed in the context of online learning, can be employed to study sample complexity of certain Reinforcement Learning (RL) problems. The project will focus on both Frequentist and Bayesian approaches to RL, focusing on function approximation, adaptivity, and randomized exploration.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- May 08, 2020
- Source ID
- N000142012336
Entities
People
- Alexander Rakhlin
Organizations
- Massachusetts Institute of Technology
- Office of Naval Research
- United States Navy