NEFT for Godot

NeuroEvolution of Fixed Topology library for Godot focused on modularity.

Projects

Here is a project that uses NEFT for godot. Made by zwometer

https://zwometer.itch.io/ai-racing

Setup - Installation

Due to the single inheritance principle in Godot, the setup requires some work but the library has been thought to be as easy as possible to implement.

This part will not teach neuroevolution. I recommend The Nature Of Code by Daniel Shiffman (chapter 9 and 10) for that.

1. Download this project.

   git clone https://github.com/leopnt/neft_godot.git
   

   Or download and extract the .zip archive.

2. Copy the folder neft_godot into your project folder.

3. Add an Organism Node as a child of the object that you want to train.

4. Add a Population Node as a child of your training scene. Specify the object (.tscn) that you want to train from the editor inspector: organism_parent_scene

5. Tweak the input_size, hidden_layers_sizes and output_size from the editor inspector as needed on the Organism

   

   So for this config you'll have:

   • 3 inputs
   • 2 hidden layers: [6, 4]
   • 1 output

6. In the object's code that you want to train, update the fitness with the methods of Organism

   • get_fitness()
   • set_fitness(new_fitness)
   • add_fitness(amount)
   
   

   example:

   func _process(delta):
       $Organism.add_fitness(0.1)
   

7. Add your own code to define when the population is ready for the next generation (every member has finished training). Then it's as simple as calling:

   $Population.next_generation()
   

   example:

   func _process(delta):
   
       var everybody_ready = true
       for child in $Population.get_children():
           if !child.is_ready():
               # at least one child is not ready
               everybody_ready = false
               break
   
       if everybody_ready:
           $Population.next_generation()
   
           # reset the world and children
           # [...]
   

8. To take decisions from the neural network of the Organism, you have to define the input and give it to the appropriate method:

   var inputs:Array = []
   inputs.push_back(x1)
   inputs.push_back(x2)
   [...]
   
   var output:Array = $Organism.think(inputs)
   
   # then you can take decisions from the output
   

   Note that the input size must match the neural network's input size.

   Also note that the output is an array.

   It's advised to normalize the inputs (values between 0 and 1) as much as possible to avoid input issues.

Demo

You can find an example in the example folder. Here is some data extracted from the Flappy Bird example. I ran it with Engine.time_scale = 2

Population size: 200
Mutation rate: 1%

As you can see, the average best score continuously increases and it slows down until it eventually reaches infinity (not visible here) after enough iterations.