Multiple array variable

This approach for a neuronal network is not very flexible. I just wrote a simple NeuronalNetwork framework which is fully scalable. You can create a neuronal network with as many hidden layers you like in any configuration with a single line of code. The only thing that's fixed is that each node of one layer is connected to each node of the next layer.

 NeuronalNetwork network = new NeuronalNetwork(inputCount, outputCount, hiddenLayer1, hiddenLayer2, ...);

Some examples:

 // 2 inputs, 1 outputs, 3 neurons in the first hidden layer, 5 in the second hidden layer and 2 in the third. 
 network = new NeuronalNetwork(2, 1, 3, 5, 2);
 
 // 4 inputs, 2 outputs, 5 neurons in the hidden layer
 network = new NeuronalNetwork(4, 2, 5);

You can set the inputs with SetInput and read the outputs with GetOutput. Alternatively you can access the input and output layers directly. You have to call "Calculate()" after you set the inputs in order to calculate the output values.

You can easily clone a NeuronalNetwork with the copy constructor:

 NeuronalNetwork clone = new NeuronalNetwork( source );

This will duplicate the whole network and copy the weights and settings over. You can call the Mutate method in order to randomly mutate the network. I implemented this the same way as in the source code you've linked.

Note: When Mutate is called with a MutationRate of "-1" it will randomize all stats. This is automatically called for new networks.

At the moment it doesn't have any kind of backpropagation or weight adjustments since the original code only depends on random evolution.

edit

I just implemented two ways of serializing a "NeuronalNetwork":

• XML

• Binary

To serialize a network as an XML string you would do this:

XML

 NeuronalNetwork network;
 
 var doc = new System.Xml.XmlDocument(); // new XML document
 var nets = doc.AppendChild(doc.CreateElement("Networks")); // add root node
 network.Serialize(nets); // serialize the network into the root node
 // the following is just converting the XML document into a string
 var sb = new System.Text.StringBuilder();
 using (var writer = new System.IO.StringWriter(sb))
     doc.Save(writer);
 string XMLtext = sb.ToString();

To deserialize you would do:

 var doc = new System.Xml.XmlDocument();
 doc.LoadXml(text);
 var nets = doc["Networks"]; // root node
 var node = nets.ChildNodes.Item(0); // get the actual network node
 NeuronalNetwork network = new NeuronalNetwork(node);

To serialize it as a byte array you would do:

Binary

 NeuronalNetwork network;
 
 byte[] data;
 using (var stream = new System.IO.MemoryStream())
 using (var writer = new System.IO.BinaryWriter(stream))
 {
     network.Serialize(writer);
     data = stream.ToArray();
 }

To deserialize:

 NeuronalNetwork network;
 byte[] data;
 
 using (var stream = new System.IO.MemoryStream(data))
 using (var reader = new System.IO.BinaryReader(stream))
 {
     network = new NeuronalNetwork(reader);
 }

Of course instead of a MemoryStream you could directly use a file stream if you want to read / write to a file.

Likewise the XmlDocument has a Load and Save method which also works with IO streams.

Not sure if this is what you're looking for:

 // Create the jagged array
 float[][] input = new float[LayerAmount][];
 for (int i = 0; i < LayerAmount; i++)
 {
     input[i] = new float[NeuronAmountForEachLayer];
 }
 
 // Randomize the values
 for (int i = 0; i < LayerAmount; i++)
 {
     for (int j = 0; j < NeuronAmountForEachLayer; j++)
     {
         input[i][j] = Random.Range(-1f, 1f);
     }
 }

They're called jagged arrays: https://msdn.microsoft.com/da-dk/library/2s05feca.aspx