neat-ml

An implementation of the NEAT (Neuroevolution through augmenting topologies) algorithm in Java. Originally found at http://nn.cs.utexas.edu/downloads/papers/stanley.ec02.pdf

Features

Fully documented code base ideal for learning and people new to the subject
- Documentation can be found here
Super friendly highly abstracted interfaces that hide the implementation details, so you can focus on your use case
- Friendly implementations of both genotypes and phenotypes
- All you need to worry about is inputs, outputs and fitness!
Multi-threading support for concurrent genotype evaluation just by specifying the number of worker threads
Neural network visualiser to create images, so you can see what networks are created
Full serializing and deserializing support using google's protobuf, allowing you to save interesting genotypes (networks) or even save your training progress to disk!

Quick start

This library has been built with the intention of making it as easy as possible to add NEAT to any project - without the need for deep understanding of how NEAT works.

Take the following quick example that shows how to use this library to learn the XOR function:

TLDR

a little bit of config
create evolution (all the hard work has been done for you!)
- Write the evaluator for a genotype
  - Convert the genotype into a neural network with the provided functions
  - Give the neural network inputs and receive its outputs in a single function
Run the evolve function once for every generation
Get the best performing genotype from the population
Create a neural network from that genotype as before
Done! You have a neural network that has learnt your desired behaviour

import java.util.Random;
import jwblangley.neat.evolution.Evolution;
import jwblangley.neat.evolution.EvolutionFactory;
import jwblangley.neat.phenotype.Network;
import jwblangley.neat.genotype.NetworkGenotype;

public class LearnXor {
  public static void learnXor() {

    Random random = new Random();

    // Some configuration for the evolution
    final int numGenerations = 100;
    final int populationSize = 100;
    final int targetNumSpecies = 5;
    final int numProcessingThreads = 8;

    // Create a new optimisation with 2 inputs and 1 output
    Evolution evolution = EvolutionFactory.createOptimisation(2, 1, populationSize,
      targetNumSpecies, numProcessingThreads, networkGenotype -> {

          // Build a neural network from the genotype. This is all done for you with this method!
          // There is an equivalent method for building linear output networks
          Network network = Network.createSigmoidOutputNetworkFromGenotype(networkGenotype);

          // Now that a network has been created, evaluate it however you see fit!
          // To prevent over-fitting to a particular problem, we want to actually evaluate it
          // several times and aggregate a score (For XOR there are only 4 possible combinations,
          // but this is a good habit to get into for the general case).
          // N.B: If you are trying to solve a non-generalised problem, this is not needed
          // e.g. solving a particular game level that is deterministic (has no random elements)
          int numCorrect = 0;
          for (int i = 0; i < 5; i++) {
            final double a = Math.round(random.nextDouble());
            final double b = Math.round(random.nextDouble());

            // This is as simple as it is to get the output from your neural network!
            // We get the 0th index item as we want the first (and in this case only) output
            final double output = network.calculateOutputs(a, b).get(0);

            // Using "> 0.5" here is a good way to turn a sigmoid output (0-1) into a binary output!
            final boolean expected = (a > 0.5) ^ (b > 0.5);
            final boolean actual = output > 0.5;

            if (expected == actual) {
              numCorrect++;
            }
          }

          // Return a score that tells the program how well the network did! Higher is better!
          return numCorrect * 100d / testsInEvaluate;
        });

    // Evolve!
    evolution.setVerbose(true);
    for (int i = 1; i <= numGenerations; i++) {
      // This call is all you need to evolve your population (one generation)!
      evolution.evolve(random);

      System.out.println();
    }

    // After the evolution is done (you will need to experiment to know how much training you need)
    // Get the best genotype from the population
    NetworkGenotype bestInPop = evolution.getFittestGenotype();

    // Create the network from this genotype as we did before
    Network bestNetwork = Network.createSigmoidOutputNetworkFromGenotype(bestInPop);

    // Congratulations! You now have a neural network that knows XOR
    boolean ff = bestNetwork.calculateOutputs(0d, 0d).get(0) > 0.5;
    boolean ft = bestNetwork.calculateOutputs(0d, 1d).get(0) > 0.5;
    boolean tf = bestNetwork.calculateOutputs(1d, 0d).get(0) > 0.5;
    boolean tt = bestNetwork.calculateOutputs(1d, 1d).get(0) > 0.5;

    System.out.println("false XOR false = " + ff);
    System.out.println("false XOR true = " + ft);
    System.out.println("true XOR false = " + tf);
    System.out.println("true XOR true = " + tt);

    // Prints:
    // false XOR false = false
    // false XOR true = true
    // true XOR false = true
    // true XOR true = false

  }
}

Visualiser

Visualising a neural network is a great way to understand a bit about what is going on and it also looks great! The visualiser in this library has the following features:

Different colours for positive and negative weights
- Orange for positive
- Purple for negative
Different thicknesses based on connection weight
- Thicker connections indicate greater (absolute) weights
Tapers to show direction
- Each connection is tapered from thicker to thinner to indicate connection direction

Visualising a network is as simple as:

// network is genotype - for example from evolution.getFittestGenotype()
BufferedImage image  = Visualiser.visualiseNetwork(network);

You are then free to handle the standard java BufferedImage as you see fit. However, if you want to write the image to you disk you can utilise this library's function:

Visualiser.saveImageToFile(image, new File("network.png"), true);

This can achieve results like this:

Serialise

Networks

You can save a particular network genotype for later use. This could be so that you can keep track of the best/most interesting individuals and have them evaluate without the need for training. This is a great utility with many uses: for example embedding in games as the AI!

To write a network genotype to a file:

ProtoIO.toFile(networkGenotype, new File("network.geno"));

To read a network genotype from a file:

NetworkGenotype fromProto = ProtoIO.networkFromFile(new File("network.geno"))

Note that these networks are genotypes, not phenotypes. You will need to create them in the same way as you did when training before you can use them:

// Or linear output!
Network network = Network.createSigmoidOutputNetworkFromGenotype(fromProto);

network.calculateOutputs(/* inputs here */);

Evolution

With this library comes the very powerful ability to stop training (evolving), save your progress and resume training again at a later date without losing progress.

Evaluator evaluator = networkGenotype -> {
  /* evaluate here */
  return /* fitness */;
};

int populationSize = 100;
int targetNumSpecies = 5;
int numThreads = 8;

Evolution evolution = EvolutionFactory.createOptimisation(
  /* numInputs*/,
  /* numOutputs */,
  populationSize,
  targetNumSpecies,
  numThreads,
  evaluator);

for (int i = 0; i < 100; i++) {
  evolution.evolve();
}

/*
  .
  .
  .
  Get interrupted here
  .
  .
  .
*/

// Write to disk
ProtoIO.toFile(evolution, new File("evolution.evo"));

/*
  .
  .
  .
  Want to resume here
  .
  .
  .
*/

Evolution loadedEvolution = ProtoIO.evolutionFromFile(
  new File("evolution.evo");
  targetNumSpecies,
  numThreads,
  evaluator
);

// Continue evolution
for (int i = 0; i < 100; i++) {
  loadedEvolution.evolve();
}

This does actually give you the ability to define a new Evaluator (and the other params passed into evolutionFromFile) midway through training. This is normally not recommended, but can actually be very useful for some incremental learning techniques.

jwblangley/neat-ml