[Colloq] Speaker: Chongjie Zhang, Tsinghua University - Towards Knowledge Transfer and Generalization in Deep Reinforcement Learning

[Ponte, Christopher]

Speaker: Chongjie Zhang, Tsinghua University
When: 10:30am August 8th
Where: ISEC 655
Host: Chris Amato

Title: Towards Knowledge Transfer and Generalization in Deep Reinforcement Learning

Abstract:

Deep reinforcement learning has recently shown considerable successes of achieving human-level control in some games. However, learning each task independently and from scratch requires vast experiences and is not tractable for most real-world problems. To address this challenge, in this talk, I will first present model-free policy reuse approaches that transfer past source policies to new tasks with theoretical optimality guarantees. I will then discuss an object-oriented dynamics learning approach for enabling generalizable model-based deep reinforcement learning. This approach enables a dynamics model learned from few environments to generalize over novel environments while providing semantically and visually interpretable representations.

Christopher Amato
Assistant Professor
College of Computer and Information Science (CCIS)
Northeastern University
http://www.ccs.neu.edu/home/camato/<https://na01.safelinks.protection.outlook.com/?url=http%3A%2F%2Fwww.ccs.neu.edu%2Fhome%2Fcamato%2F&data=02%7C01%7C%7C79b0323395e744802cb508d5f632de8d%7Ca8eec281aaa34daeac9b9a398b9215e7%7C0%7C0%7C636685620686175461&sdata=3HnUpV8%2FIxcUHPUvg5rMcVRM006ExF9pdOUAijgdpWo%3D&reserved=0>