One of the most exciting potential benefits of big data is that it will enable better decision making in applications such as teaching software, patient treatment plans, and user personalization. In many complex domains involving people, the number of descriptions of a situation and the number of potential decisions that can be made to improve upon the previous situation is enormous. The sheer size of this problem implies that careful selections of additional data must be made to improve decision making. In health, these decisions could involve a course of treatment; in learning these could involve remediations such as additional practice, in personalization it could involve improved choices for consumers. This proposal will develop new algorithms that are especially effective in coming to close to optimal decisions quickly.

This project will aim to develop novel algorithms that specifically address these needs, and thus enable data intensive systems to be efficiently used towards solving the decision making process. Since novel algorithmic methods of this kind are best developed, refined and tested on a real domain, the project will focus on the educational domain of classroom learning activities that optimize learning outcomes. However, the algorithms will be applicable to numerous other decision making domains. Specifically the project will focus on finding algorithmic solutions for:

(1) Large-scale deployment and data collection of lesson enactments in thousands of classrooms worldwide. (2) Data-driven methods to efficiently determine the likely most important parameters of the classroom decision making process. (3) The development of novel reinforcement learning algorithms with a focus on efficient use of data to rapidly converge towards beneficial policies. (4) Using data to determine the key model and decision parameter space bottlenecks, which, if removed, could significantly improve the outcomes of the Reinforcement Learning process. The efficacy of the optimized policy will be determined by the improvement of classroom learning performance at scale.

The methods that will be developed by this project are general enough to be directly applicable to domains of patient treatment, or any other domain that involves unknown model dynamics and decision space that involves people. The Reinforcement Learning methods will similarly find practical use in situations where a decision policy is deployed in a large number of instances asynchronously and in other high-risk settings where reducing over-exploration as much as possible is of high importance.

National Science Foundation (NSF)
Division of Graduate Education (DGE)
Standard Grant (Standard)
Application #
Program Officer
John Cherniavsky
Project Start
Project End
Budget Start
Budget End
Support Year
Fiscal Year
Total Cost
Indirect Cost
University of Washington
United States
Zip Code