Bubbles 9000
Bubbles9000 is a very simple game developed as learning aid and to serve as a test platform in exploring some basic AI / machine learning concepts. The game was developed using Processing. Aim of the game is to shoot the bubbles that float around on screen and not die. You can also shoot other players for points. There can be 1-N players depending on configuration.
Starting the Game
On windows run the application ./Bubbles9000/application.windows32/Bubbles9000.exe
On Linux run the application ./Bubbles9000/application.linux64/Bubbles9000
No other system supported.
Rules Summary
- +100 points for shooting and popping a bubble.
- +500 points for shooting and killing another ship.
- Score set to -500 if killed.
- All ships start with +100 health. Ship is considered killed if health is 0.
- Ships are bound inside a 2D world and bounce of walls, bubbles and each other.
- Bumping into a bubble or ship causes both to loose health. The angle of attack and velocity effects how much health is reduced.
Ship Controls
There are four controls of a ship:
- Steer left.
- Steer right.
- Thrust forward.
- Shoot.
There is no limit to how much you can shoot, and no penalty for shooting continuously (there probably should be. Todo in v2).
Autonomous Ship Inputs
Autonomous ships are feed 4 input which they use to decide which ship controls to activate.
- Distance to the nearest ship.
- Angle to the nearest ship.
- Distance to the nearest bubble.
- Angle to nearest bubble.
Angles range from [-PI+PI/8,PI-PI/8] and is 0 when the target is directly in front.
Autonomous Ship Training
Two methods were used.
- Offline evolutionary method: A number of ships compete over a number of iterations to build up a pool of <config, score> pairs. These pairs are then evolved over a number of generations.
- Online supervised: The game is start with one human player and another computer player. The computer player trains it's neural net given the live input from the human player.
Neural Network Used
I just used a perceptron network with 4 inputs and 4 outputs.
Source Code Overview
- Bubbles9000/ Processing application including pre trained neuron configurations, and a build of the final application for Window and Linux.
- ShipEvolver/ Scripts used to evolve neuron configurations offline.
- README.md This file.
- pres.mp4 Video presentation.