Electromechanical Platform with Removable Overlay for Exploring, Tuning and Evaluating Reinforcement Learning Algorithms

Abstract

Presented a physical electromechanical movable maze platform for evaluating reinforcement learning (RL) algorithms. The use of embedded hall sensors in the platform for detecting the spherical magnetic ball provides benefits over top mounted camera systems. The process of adapting RL algorithms like Q-table, SARSA and Neural Network to function with the platform was discussed. A comparative evaluation of the performance against baseline was presented. The electromechanical platform provides unique features, benefits, and challenges. The platform serves as a tool in RL algorithm tuning and validation. The platform also serves as a pedagogical tool, especially in providing learners a means to visualize the RL algorithms in action.