Loading
Preliminary Exam Talks
Using Images As State in Reinforcement Learning
|
Abstract
Reinforcement learning is a powerful technique in which an agent acting in a world can learn which actions to take so as to maximize its expected future rewards. However, RL techniques do not work well for very high dimensional input data, such as that obtained from images. We propose a framework by which images can be reduced to a lower dimensional state space in which RL techniques can operate, and a way of projecting images into this low dimensional space so agents can execute the learned policies on new images.
Advisor(s): Carlo Tomasi
Ronald Parr, Guillermo Sapiro, Mauro Maggioni