Deep Reinforcement Learning with Mario Games

Apply a DQN model to play Mario games by Shiqi Zheng. Final Project for CIS4930 Deep Learning for Computer Graphics Fall 2020 instructed by Dr. Corey Toler-Franklin at University of Florida.

Introduction

Demo

DQN Super Mario Player Structure Overview

To read more details, please view Deep Reinforcement Learning with Mario Games.pdf.

Setup

Install python and jupyter environment

conda create -n DQN_SuperMario python=3.7 jupyter   # You can also use other environment.

Install other dependencies

pip install -r requirements.txt

Training

Use jupter notebook to run DQN_super_mario.ipynb.
It takes a while to train, and if you do not want to wait, Local_DQN_Mario_big_4 is a trained model that can be used to play the game.

Testing

Use jupter notebook to run Mario_Play.ipynb with the model created in the Training.

Result

The 100 times rewards results are summarized in Table 1.

The last 100 reward trend is shown in Figure 2.

Acknowledgements

• Part of the code is built based on Playing Atari with Deep Reinforcement Learning by Mnih, Volodymyr, et al.
• Also thanks to John N Tsitsiklis and Benjamin Van Roy and their paper An analysis of temporal-difference learning with function approximation. Automatic Control, IEEE Transactions on, 42(5):674–690, 1997.
• Thanks to Deep-Reinforcement-Learning-Hands-On and REINFORCEMENT LEARNING (DQN) TUTORIAL.