Experimental Evaluation of Reinforcement Learning Algorithms

Abstract

Reinforcement learning is an active field of machine learning that deals with developing agents that take actions in an environment with the end goal of maximizing the total reward. The field of reinforcement learning has gained increasing interest in recent years, and efforts to improve the algorithms have grown substantially. To aid in the development of better algorithms, this paper tries to evaluate the state-of-the-art reinforcement learning algorithms for solving the task of learning with raw pixels of an image as input to the algorithm by testing their performance on several benchmarks from the OpenAI Gym suite of games. This paper compares their learning capabilities and consistency throughout the multiple runs and analyzes the results of testing these algorithms to provide insights into the flaws of certain algorithms.