flappyNetwork

A neural network learning how to play flappy bird

How to use

You can start the compiled file without any arguments to start training a new network, that will learn how to play flappy bird.

Alternatively you can specify to play against an already trained network, by specifying its name. E.g.

main TRAINED.gob

Controls

When playing the actual game, there is a white square and a black one. The white square is controlled by the best performing network or the one you loaded at the beginning. The black square is controlled by you.

• Space is used to make your bird jump
• R is used to reset the game
• G will make your bird invulnerable/revive it
• ArrowLeft slows down the time between frames
• ArrowRight speeds up the time between frames

Debug Information

In the top right are the neural networks inputs and its output displayed. Underneath them you can find the TPS and FPS, aswell as artificially introduced time between frames (the slow motion controlled via the arrow keys)

Neural Network

The neural network that is used is an evolutionary neural network.

Structure

The neural network has 6 inputs and 1 output, that being a float between 0.0 and 1.0 (true/false)

Each frame/tick the neural network gets these inputs:

• gSpeed, the current speed of the game
• bHeight, the height of the bird
• bVel, the current (falling) speed of the bird
• pDst, the distance to the next pipe
• pU, the coordinates of the upper pipe
• pL, the coodinates of the lower pipe

Based on these inputs, the network returns wether or not it wants to press jump on the current frame. If the output is > 0.5 the network jumps.

Evolution

To start, {population size} networks are created with random weights and biases.

Based on the fitness function, in this case how far the bird controlled by the network came, the network gets an amount of points. The best networks make it to the next generation without any modifications (elitism/"survival of the fittest"). Half of the rest gets crossed over with the previously mentioned best networks (procreation). The rest gets randomly mutated based on a mutation score

This process is repeated until the desired amount of generations is met or until 50.000 points are scored by a network, which signals the network has evolved enough

All of this means that the programm is able to train a network to play flappy bird without any help from the outside, simply from knowing which network performed good and which one did not

Purpose

This doesn't really serve any purpose, except trying out a neural network i built from scratch. This code was mostly written late at night so don't judge it too much

Disclaimer

While this might run on linux, i've only tested it on windows.