Title: Deep Reinforcement Learning with Prior Knowledge

Abstract:

Deep reinforcement learning has been applied to many domains from computer games, natural language processing, recommendation systems to robotics.  While model-free reinforcement learning algorithms are promising approaches to learning policies without knowledge of the system dynamics, they usually require much more data. In this thesis, we examine the importance of using prior knowledge in deep reinforcement learning algorithms, especially in model-free settings. We show that applying prior knowledge can significantly improve the learning speed and generalization capabilities. More specifically, we discuss the applications of inducing prior knowledge in two problems: multi-robot transfer learning, and exploration in navigation. In multi-robot transfer learning, we found that using domain randomization as well as kinematics encoding is able to train a policy that generalizes to new robots with varying degree-of-freedoms, kinematics, and dynamics. In learning exploration policies for navigation, we show that using occupancy map, obtained by projecting 3D reconstructed map, can serve as a spatial memory for the policies and the policies can retain its exploration abilities in new houses.

Committee:

Abhinav Gupta (advisor)

Oliver Kroemer

Adithyavairavan Murali