T-Rex Runner - Neural Network AI

Introduction

This project's goal is to create an AI to beat Chromium's T-Rex Runner game.

The needed variables for the AI are :

• The current T-Rex's speed
• The distance to the nearest obstacle
• The width of the nearest obstacle.

(At first, there will be no pterodactyl management. This could be fixed by getting the height of the next obstacle)

And to evaluate an IA's performance:

• The number of passed obstacles.

(As the obstacle apparition rate is random, the ran distance is not a good evaluation criter)

Those variables are displayed in real time on the debugged T-Rex Runner game (neterror.debug.js), below the score:

Coding the AI

1. The naive way

The simplest way to proceed is to make the dino jump when the nearest obstacle is clos(about 150px away). This works, but not for long : I have not yet managed to reach a score over 1k using this method.

2. The analytic way

Another possibility is to find an application taking in parameters the speed, dist, width, and returning 1 (UP) or 0 (nothing), i.e

f: R^3 -> {0, 1}

The problem is in determining the parameters' coefficient, and the function type (linear? polynomial?)

3. Genetic Algorithm

See AI/genetic.py

To determine such coefficients, a genetic algorithm (GA) could be a good solution. We fix the fuction type as multilinear, i.e.

f(s, d, w) = a0*s + a1*d + a2*w)

Here is a screenshot of the GA running:

4. Neural Networks

Yet to come

Required dependencies

In order to run the AI files, the following python packages are required :

• numpy
• pyautogui
• virtualenv
• selenium

And selenium needs its own Chrome/Chromium driver :

Download the latest version here : https://sites.google.com/a/chromium.org/chromedriver/downloads

Then run: unzip path/to/chromedriver_XXX.zip && sudo mv path/to/chromedriver /usr/local/bin/chromedriver