Gym Hero: A Research Environment for Reinforcement Learning Agents in Rhythm Games

This work presents a Reinforcement Learning environment, called Gym Hero, based on the game Guitar Hero. It consists of a similar game implementation, developed using the graphics engine PyGame, with four difficulty levels, and able to randomly generate tracks. On top of the game, we implemented a Gym environment to train and evaluate Reinforcement Learning agents. In order to assess the environment's capacity as a suitable learning tool, we ran a set of experiments to train three autonomous agents using Deep Reinforcement Learning. Each agent was trained on a different level using Deep Q-Networks, a technique that combines Reinforcement Learning with Deep Neural Networks. The input of the network is only the pixels of the screen. We show that the agents were capable of learning the expected behaviors to play the game. The obtained results validate the proposed environment as capable of evaluating autonomous agents on Reinforcement Learning tasks.