HooIsNeural

A Java implementation of Backpropagation, Feed-Forward, and NEAT networks.

Screenshots

Examples

Backpropagation Network

Creating:

Use the network builder class to generate a Backpropagating network.

(BackPropNetwork) NetworkBuilder
	.create(NetworkType.BACKPROP)
	.setLabel("XOR GATE")
	.setInputLayer(
			LayerBuilder
			.create()
			.addNodes(2, Activations.sigmoid))
	.addHiddenLayer(
			LayerBuilder
			.create()
			.addNodes(4, Activations.sigmoid))
	.setOutputLayer(
			LayerBuilder
			.create()
			.addNodes(1, Activations.sigmoid)
			)
	.withBiasNode()
	.withVisualizer()
	.withSettings(
		BackPropSettings
		.create()
		.setLearningRate(0.04)
		.setRandomSeed(3000))
	.createAllConnections()
	.build();	

Training:

To train the network execute the following code.

private static final double MIN_ERROR = 0.0001;
private static final int TRAINING_LIMIT = 250000;

int cycle = 0;
while (network.getTotalError() > MIN_ERROR && cycle < TRAINING_LIMIT) {
	network.clear();
	network.train();
	cycle++;
}

Executing:

To get the output execute the following code.

network.clear();
network.setInput(input);
network.calculate();
network.getOutput();

NEAT Network

Creating:

Using the network builder class we can also create a NEAT network.

(NeatNetwork) NetworkBuilder
	.create(NetworkType.EVOLUTION)
	.setLabel("XOR GATE")
	.setInputLayer(
			LayerBuilder
			.create()
			.addNodes(2, Activations.sigmoid))
	.addHiddenLayer(						
			LayerBuilder
			.create()
			.addNodes(0, Activations.sigmoid))
	.setOutputLayer(
			LayerBuilder
			.create()
			.addNodes(1, Activations.sigmoid)
			)
	.withBiasNode()
	.withSettings(
		NeatSettings
		.create()
		.setRandomSeed(3000)
		.setHiddenNodeCap(4)
		.setHiddenLayerCap(1)
		.setMaxMutations(5)
		.setWeightAdjustment(1)
		.setMaxWeight(10.0)
		.setMinWeight(-10.0)
		.setPopulationSize(500)
		.setActivation(Activations.sigmoid))
	.build();	

Training:

First we implement the NeatFitness class to determine the fitness of each network during the evolution process. In this case, it is the XOR function.

public class XORFitness implements NeatFitness {

	@Override
	public double calculate(Network network) {
		double[][] inputs = new double[][] {{1, 1}, {1, 0}, {0, 1}, {0, 0}};
		double[][] outputs = new double[][] {{0}, {1}, {1}, {0}};

		double score = 0;
		for (int i = 0; i < inputs.length; i++) {
			double[] input = inputs[i];
			double[] output = outputs[i];
			network.clear();
			network.setInput(input);
			network.calculate();

			double[] actual = network.getOutput();
			double err = Math.abs(output[0] - actual[0]);
			score += output[0] == 0 ? 1.0 - err : 1.0 - (err * 2);
		}

		return score;
	}
}

Next, we apply the fitness function to the network and we begin the evolution process.

network.setFitness(new XORFitness());
int expectedFitness = 3.5;
int generation = 0;
while (network.getFitness() < expectedFitness) {
	network = network.train();
	generation++;
}

Executing:

We first clear any existing input and then set ours, and execute the feed-forward process to get the output.

network.clear();
network.setInput(new double[] {1,1});
network.calculate();
network.getOutput();

Requirements

• Java 8+
• Maven 1.3+