Analysis of Reinforcement Learning in Maze Environment

Savita Kumari Sheoran, Poonam

Abstract Maze environment presents complex path grids made of an arbitrary number of squares of varying width and length with restricted movements. Since its geometry incorporate varying level of knowledge, hence presents interesting challenge before Artificial Intelligence (AI) community. Reinforcement Learning is used to trace optimal solution in maze learning environment, where agent learns its behavior through trial and error. The discrete Q-Learning, Dyna-CA and Fuzzy Rule Interpolation-based Q-learning (FRIQ-learning) are commonly used and proven Reinforcement Learning methods in Machine Learning to solve such puzzles. This research paper aim to finding out such method which converges in minimum time. We he simulated the maze environment with and without obstacles configurations over MATLAB computational platform to compare the real time parameter of convergence time. The performance results were analyzed and presented. The final results reveals that FRIQ-learning outperform the others under all conditions. Keywords: Reinforcement learning, discrete Q-learning, DYNA-CA learning, FRIQ-learning, maze problem.

Title:	Analysis of Reinforcement Learning in Maze Environment
Author Name:	Savita Kumari Sheoran, Poonam
Abstract:	Abstract Maze environment presents complex path grids made of an arbitrary number of squares of varying width and length with restricted movements. Since its geometry incorporate varying level of knowledge, hence presents interesting challenge before Artificial Intelligence (AI) community. Reinforcement Learning is used to trace optimal solution in maze learning environment, where agent learns its behavior through trial and error. The discrete Q-Learning, Dyna-CA and Fuzzy Rule Interpolation-based Q-learning (FRIQ-learning) are commonly used and proven Reinforcement Learning methods in Machine Learning to solve such puzzles. This research paper aim to finding out such method which converges in minimum time. We he simulated the maze environment with and without obstacles configurations over MATLAB computational platform to compare the real time parameter of convergence time. The performance results were analyzed and presented. The final results reveals that FRIQ-learning outperform the others under all conditions. Keywords: Reinforcement learning, discrete Q-learning, DYNA-CA learning, FRIQ-learning, maze problem.
Cite this article:	Savita Kumari Sheoran, Poonam , " Analysis of Reinforcement Learning in Maze Environment " , International Journal of Emerging Trends & Technology in Computer Science (IJETTCS), Volume 6, Issue 1, January - February 2017 , pp. 021-025 , ISSN 2278-6856.
Full Text [PDF] Home