zefman / Brainwave

A JavaScript neural network learning using genetic algorithms

Brainwave

A JavaScript neural network learning using genetic algorithms. To be more specific, a Javascript implementation of the Neural network described here on the AI Junkie site: http://www.ai-junkie.com/ann/evolved/nnt1.html

You can find examples of it working here: http://jozefmaxted.co.uk/neural/index.html

Using Brainwave

To begin simply include Brainwave.js on your page

<script src="build/Brainwave.js"></script>

Brainwave has two main components, Network which of course is the neural net, and Genetics which is the genetic algorithm used to improve a population of networks.

##Creating a Neural Network

A network of varying size and structure can be created easily with the Network object. When a new network is first created its weights and biases are all initialised randomly.

var network = new Brainwave.Network(numInputs, numOutputs, numHiddenLayers, numNeuronsPerHiddenLayer);

###Running the network The network expects an array of floats to be passed as inputs. After being passed through the network, an array of output values will then be returned.

var network = new Brainwave.Network(4, 2, 1, 4);

network.run([3.56, 2.1, 18.9, -4.7]);

// Possible output
// [ 0.656, 0.983 ]

Evolving A Population of Networks

I plan to add a training helper object to make the process of setting up a bunch of networks and training them super simple once I get time, but currently networks can still be trained without too much trouble.

var popSize = 20;

// Create an array to hold the population of networks
var networks = [];

// Populate the networks array
for (var i = 0; i < popSize; i++) {
    networks.push(new Brainwave.Network(4, 1, 1, 4));
}

// Next we need to create the Genetics object that will evolve the networks for us
var genetics = new Brainwave.Genetics(popSize, networks[0].getNumWeights());

// When creating the genetics object it will also generate random weights an baises for the networks
// These should be imported into the population of networks before beginning any training
for (var j = 0; j < popSize; j++) {
    networks[j].importWeights(genetics.population[j].weights);
}

// Now the networks and genetics are all set up training can begin. Pass each network an input and issue
// it a fitness depending on how close its output was to the desired output
for (var k = 0; k < popSize; k++) {
    var output = networks[k].run([1, 4, 6, 2]);

    // Lets just suppose we are looking for an output as close to one as possible
    var fitness = 1 - Math.abs(1 - output); // So the max fitness we can have here is 1

    // Now we need to update the genetics with this fitness
    genetics.population[k].fitness = fitness;
}

// After you have decided on a fitness for each network, we can use the genetics
// object to evolve them based on the results
genetics.epoch(genetics.population);

// Then we just need to import the new weights into the networks and repeat again and again
for (var n = 0; n < popSize; n++) {
    networks[n].importWeights(genetics.population[n].weights);
}