Optimized Path Planning in Reinforcement Learning
by Backtracking
Volume 1 - Issue 4
Morteza Kiadi1, Qing Tan2* and José R Villar1
-
Author Information
Open or Close
- 1Computer Science Department, University of Oviedo, Spain
- 2School of Computing and Information Systems, Athabasca University, Athabasca, Canada
*Corresponding author:
Qing Tan, School of Computing and Information Systems, Athabasca University, 1 University Drive,
Athabasca, AB, T9S 3A3, Canada
Received: June 28, 2019; Published: July 26, 2019
DOI:
10.32474/CTCSA.2018.01.000116
Full Text
PDF
To view the Full Article Peer-reviewed Article PDF
Abstract
Finding the shortest path to reach a destination in an indoor environment is so important that it is deemed as one of the
fundamental and classical topics of Reinforcement Learning. Identifying the best path to reach a destination in a maze has been
used as a testbed to learn and simulate Reinforcement Learning algorithms. In this paper, we developed a faster pathfinding model
by optimization of the Markov decision-making process and tweaking the model’s hyper-parameters. The proposed method allows
researchers to use the suggested optimized model in bigger state spaces. The presented results assist the researchers in this field to
get a better idea of the Reinforcement Learning tenets and contributes to Reinforcement Learning community by making this topic
more accessible. The problem we are going to solve it in the context of finding the shortest path in a smart lab “as soon as possible”.
In this paper, we prove that the shortest path in a maze appears much sooner than finalizing the state values. By optimizing the code
to behave smarter and extracting the recurrent visited states, we accelerate the search process to the magnitude of three times.
Moreover, we compare the execution of the optimized codes on the dedicated cloud servers to simulate offloading the processing
power from a robot to a more powerful processing server but in the proximity of the robot. This matter relates to the outcome of
this paper to the Fog Computing in the context of robot pathfinding in indoor environments.
Keywords: Robot Path Planning; Reinforcement Learning; explore vs. exploit; Markov Decision Process; Smart Lab; Fog Computing
Abstract|
Introduction|
Related Work|
The Solution Components|
Exploit VS. Explore|
Hyper-Parameters and Results|
Optimization by Backtracking|
Discussion|
Conclusion|
Acknowledgements|
References|