'Practice' for Enhancing the Performance of a Deep Reinforcement Learning Agent

1 online resource (67 pages) : PDF

2019

University of North Carolina at Charlotte

Deep reinforcement learning has demonstrated its capability to solve a diverse array of challenging problems, which were not able to solve previously. It has been able to achieve human-level performance in Atari 2600 games and it has shown great potential to self-driving cars, robotics, and natural language processing. However, it requires long training time to learn feature representations from raw sensory data and makes it difficult to use deep reinforcement learning in real-world applications. Transfer learning, specifically sim-to-real transfer, has been suggested to learn features but designing of new environments or defining an environment with similar or relevant goals requires careful human efforts. This thesis aims to leverage practice approaches, which do not require a new environmental design for transfer learning, to reduce the time for training a deep reinforcement learning agent and enhance its performance. First, this thesis introduces the use of practice approach applicable to end-to-end models with a deep reinforcement learning algorithm. It shows that the approach improves the performance of an agent and presents experimental results of this approach in complex environments. Second, this thesis presents a novel strategy, iterative practice, which repeats short period of practice and short period of learning until desired results are achieved. It presents experimental results and verifies that iterative practice improves learning. Last, it introduces a new strategy, shared experience for iterative practice, which reduces the interactions with an environment. It presents observations on this approach applied in complex environments and examines the effect of reduced interactions on learning.

masters theses

Artificial intelligence

M.S.

Deep Learning
Machine Learning
Reinforcement Learning
Transfer Learning

Computer Science

Lee, Minwoo

Akella, Srinivas
Shin, Min

Thesis (M.S.)--University of North Carolina at Charlotte, 2019.

This Item is protected by copyright and/or related rights. You are free to use this Item in any way that is permitted by the copyright and related rights legislation that applies to your use. For other uses you need to obtain permission from the rights-holder(s). For additional information, see http://rightsstatements.org/page/InC/1.0/.

Copyright is held by the author unless otherwise indicated.

Kancharla_uncc_0694N_12150

http://hdl.handle.net/20.500.13093/etd:1988