M-net: A Convolutional Neural Network Based in Reinforcement Learning to Play Mario Bros

Resumen

Nowadays, Deep Reinforcement Learning (DRL) has made great progress since it was first proposed. Generally, DRL-based models take large amounts of data and make decisions according to the policy established by these models. This learning method updates the policy and maximizes the reward so that the agent can perform a task optimally. In this work, an agent is trained in order to learn how to play video games. Deep Q-learning (DQL) methods that explain the behavior and decision-making generated by the agent will be discussed. First, the agent is trained through a convolutional model to understand the environment it is learning. From this understanding, the agent makes decisions in order to maximize the reward generated by the model. Finally, the agent is put to the test in the Super Mario Bros environment in order to complete the level. The results of this work show how the agent is able to learn in the environment maximizing the reward.