Publication Date

Fall 2023

Degree Type

Thesis

Degree Name

Master of Science (MS)

Department

Computer Engineering

Advisor

Stas Tiomkin; Carlos Rojas; Mahima Agumbe Suresh

Abstract

Our work introduces a way to learn an optimal reinforcement learning agent accompanied by intrinsic properties of the environment. The extracted properties helps the agent to extrapolate the learning to unseen states efficiently. Out of all the various types of properties, we are intrigued towards equivariant and invariant properties, which essentially translates to symmetry. Contrary to many approaches, we do not assume the symmetry, rather learn them, making the approach agnostic to the environment and the property. The learned properties offers multiple perspective of the environment to exploit it to benefit decision making while interacting with the environment. By building this prior group information into our learning agent’s architecture, we essentially reduce the sample or State-action space making the interaction much more efficient and intelligent. The results are agnostic of group structures and similar results can be achieved for any environment and its related group structures. We have shown that inducing such architecture, we can learn much faster and efficiently in multiple environments using a traditional Deep-Q Network. The future scope would be to test our method across all reinforcement learning approaches, with finite and infinite State-action pairs coupled with enabling multiple properties assisting the agent. Index Terms—Group Theory, Reinforcement Learning Learning, Equivariance, Invariance.

Download

Share

COinS