Comparative Analysis of Reinforcement Learning Methods for Optimal Solution of Maze Problems

Abstract

Reinforcement learning is popular machine learning techniques for optimal planning in complex environment. The maze is a complex environment which has a grid made of an arbitrary number of squares of width and length where finding optimal path, which converge in minimum time, is always a challenging task. There are various reinforcement learning methods where agent learn from environment to find optimal path in maze problems viz. discrete Q-Learning, Dyna-CA Learning and FRIQ-Learning (Fuzzy Rule Interpolation-based Q-Learning). This research intends to carry out a comparative study of these three methods to locate a method with best convergence time. The algorithms pertaining to these methods are tested on MATLAB computational platform for different obstacles configurations of maze to compare their real time parameter of convergence time. The performance results were analyzed and presented. The final result reveals that FRIQ-Learning outperforms the others under all conditions. Keywords: Reinforcement learning, maze environment, Q-Learning, Dyna-CA Learning and FRIQ-Learning.