This thesis presents a method to find a series of actions that optimizes a series of rewards. This is with the assumption that the rewards are only known periodically after a series of actions are taken, and that there is a time dependency between actions and rewards. This setting is motivated by an additive manufacturing problem where we first create an object (make actions), and then measure its properties (observe rewards). The method makes use of Contextual Gaussian Processes to make efficient and informative predictions from past training data. The method is shown to work on synthetic data and is compared to four other algorithms designed to solve the same problem. A greedy variation is described which performs much faster than the full version and has close to optimal performance. Finally, the method is applied to a COVID dataset to predict a sequence of COVID deaths given a sequence of COVID cases. The algorithm presented in this thesis is applicable in many other fields, and is capable of finding a quicker and better optimum than similar methods

27 pages

Thesis Committee:
Aarti Singh (Chair)
Jeff Schneider

Srinivasan Seshan, Head, Computer Science Department
Martial Hebert, Dean, School of Computer Science

Return to: SCS Technical Report Collection
School of Computer Science

This page maintained by reports@cs.cmu.edu