Reinforcement Learning Applied to Doodle Jump

Overview

This project demonstrates the application of Q Learning, a reinforcement learning algorithm, to an HTML5 implementation of Doodle Jump.

Performance of the algorithm with fine-tuned hyperparameters shown below:

See paper.pdf for in-depth details.

Setup & Running

No installation is necessary, simply clone the repository and open up index.html in your favorite browser.

UX Enhancements

We've added several enhancements to the user interface for your convenience.

• Gamespeed - controls how fast the doodle plays
• Draw toggle - turn this off and the algorithm will consume less resources since it will not render anything
• Chart 1 - shows the average score per 10 deaths
• Chart 2 - shows the states explored per 10 deaths

Advice

Game

In order to modify the game's logic, see scripts/GameLogic.js

Algorithm & Hyperparameters

To modify our implementation of Q Learning, see scripts/QLearning.js

Autosave/load

Your "brain" object is automatically saved and loaded when the browser closes and opens.

Useful Variables

• brain - holds all information needed by the ML algorithm
• player - the player object holds information such as the doodle's coordinates
• platforms - holds platform information for all platforms (each of which belongs to the platform class
• gamespeed - controls how fast the doodle plays

Algorithmic Details

Please see paper.pdf for technical details of our machine learning algorithm.

Contributors

Elliot Shohet, Michael Mathew, Grigor Ambartsumyan, Anmol Singh, Ibrahim Ahmed, Tuan Vu, Haoyao Chen, Inka Arifin, Guranjan Singh, Ronil Singh, Masayuki Takagi, Ahsan Abdullah