Watch cars learn right in your browser! https://manassarpatwar.github.io/WatchCarsLearn/

• Each car has 5 distance sensors which serve as inputs to its brain (neural network)
• Each sensor has a max length and the input passed to the network is a normalized value of the sensor input (length/maxLength)
• The brain of the car outputs 2 values, both between 0 and 1
• The first value controls the throttle and brake of the car. If the value is greater than 0.66, the car accelerates else if it is less than 0.33, it brakes.
• The second value controls the steering of the car. If the value is greater than 0.66, the car steers right, else if it is less than 0.33, it steers left.
• The physics of the car is inspired by spacejack's implementation of Marco Monster's Car Physics for games paper

• I followed the steps outlined in a Procedural Race Track Generation algorithm implemented as a tool in Unity by Ian Hudson
• I used voronoi.js for creating the voronoi diagram

• I used bezier.js for working with beziers
• The track is made up of a lot of bezier curves
• There exists a hidden path of bezier curves that runs through the middle of the track. I use this path to calculate score of each car, and check whether the car is on the track or it has crashed
• The input sensor length is measured on every physics step. It is measured by calculating the intersection point between the sensor line and a nearest bezier curve

• The track is completely editable and randomizable on the fly
• The simulation is paused while the user is editing otherwise the window will freeze up
• To Edit, simply click on the pencil icon and drag any of the control points

• The network is a feed forward neural network and stops any recurrent connections from being mutated
• The network is activated using a list of nodes sorted topologically(Kahn's algorithm)

• Brightness of each node is the output of that node when activated

• Connections which are green have a positive weight
• Connections which are blue have a negative weight
• Connections which are red are disabled

• I created my own NEAT implementation in JS
• Alenaksu's implementation of NEAT in typescript helped me a lot in debugging my own
• Kenneth O. Stanley's original NEAT paper

p5.js editor sketch for pencil

p5.js editor sketch for randomize

p5.js editor sketch for steering wheel

MIT
Copyright (c) 2020-present Manas Sarpatwar