NeuralNet

This javascript module is made for applications that want to use neural network and train them using a evolutionary approach instead of backpropagation.

How to use

var neuralNet = new NeuralNet(); //This creates a random network

The NeuralNet constructor accepts two objects, the first being a settings object the second an array that represents a network. The following methods can be used:

neuralNet.mutateNetwork(settings, network);

If no mutationRate or network is provided it will use the network of the object the method was called on and the mutationRate in neuralNet.settings.mutationRate.

neuralNet.randomNetwork(hiddenLayers, nodePerLayer, inputNodes, outputNodes);

Function will return a random network. This function also defaults to the settings object.

neuralNet.runNetwork(inputNodes, network);

This function will return the calculated outputNodes, defaults to using the network of the neuralNet object it was called on.

Settings

The settings object defaults to:

{
    hiddenLayers: 2,            // The amount of hidden layers the network should have
    nodesPerLayer: 20,          // The amount of nodes for each hidden layer
    inputNodes: 2,              // The amount of input nodes the network will have
    outputNodes: 2              // The amount of output nodes
    mutationRate: 0.04          // The weights and bias will on average change this much
                                // weight += (Math.random() * 2 - 1) * mutationRate
}

Network

The internal network is an array of layer arrays:

[[], ...]

The layer arrays contain node objects

[{}, ...]

The node object contains a bias float and a weights array

{bias: 1.00000, weights: []}

The weights is an array of floats, the key represents the previous node they are connected with

[1.0000, 0.500, 0.750012168]

Everything togheter

[[{bias: 1.00000, weights: [0.0005, ...]}, ...], ...]