/// <summary>
/// Load neural network from file
/// (Recreates and retrain if settings are different. Releases training set.)
/// </summary>
public MyNeuralNetwork()
{
StopLearning();
m_curves = new Dictionary<string, ClassicCurve>();
m_network = FileOptions.LoadNeuralNetwork();
//m_curves = FileOptions.LoadTrainingSet(true);
}
NeuralNetwork.Synapsis[] extendArray(NeuralNetwork.Network nn, int i, int nc)
{
var old = nn.Synapsis[i];
nn.Synapsis[i] = new NeuralNetwork.Synapsis[nc];
old.CopyTo(nn.Synapsis[i], 0);
return(nn.Synapsis[i]);
}
private void UpdateNetwork()
{
var inputLayer = grid.Rows * grid.Columns;
var hiddenLayer = Int32.Parse(hiddenLayerTextBox.Text);
const int outputLayer = 10;
network = new Network(new[] {inputLayer, hiddenLayer, outputLayer});
}
public void ExportGeneration()
{
NeuralNetwork.Network[] generationPopulation = new NeuralNetwork.Network[networkPopulation.Count];
for (int i = 0; i < networkPopulation.Count; i++)
{
generationPopulation[i] = networkPopulation[i].network;
}
NetworkExporter.ExportGeneration(generationPopulation, generation);
}
// Use this for initialization
void Start()
{
foreach (Transform child in transform)
{
arms.Add(child);
}
ImageInput.width = ImageWidth;
ImageInput.height = ImageHeight;
int numInputs = ImageWidth * ImageHeight * 4;
int numOutputs = 3 * arms.Count;
network = new NeuralNetwork.Network(new int[] { numInputs, 12, numOutputs });
}
// Update is called once per frame
protected void Update()
{
if (network == null)
{
network = GetComponentInChildren <NeuralNetwork.Network>();
return;
}
network.Compute();
onSimulationStep.Invoke();
if (!gameOver)
{
step++;
}
}
NeuralNetwork.Network CreateNewNetwork()
{
//Create a new
NeuralNetwork.Network newNetwork = Instantiate(neuralNetworkPrefab, transform);
newNetwork.inputLayer = input;
newNetwork.onOutput.AddListener(output.SetFly);
newNetwork.InitializeConnections();
//Add the new network to the list of networks
NetworkIndividual newIndividual = new NetworkIndividual();
newIndividual.network = newNetwork;
networkPopulation.Add(newIndividual);
return(newNetwork);
}
static void Main(string[] args)
{
var mnistDataParser = new MnistDataParser();
var trainingRecords = mnistDataParser.Parse("d:\\train-images", "d:\\train-labels");
var testRecords = mnistDataParser.Parse("d:\\test-images", "d:\\test-labels");
var neuralNetwork = new Network(new[] {784, 30, 10});
var benchmark = new Benchmark(neuralNetwork);
for (double accuracy = 0; accuracy < .9;)
{
accuracy = benchmark.AccuracyFor(testRecords);
Console.WriteLine("Success rate: {0} %", accuracy*100);
neuralNetwork.Train(trainingRecords, .01, 10);
}
}
public IEnumerator RunSimulation(NeuralNetwork.Network neuralNetwork, float deltaTime)
{
step = 0;
gameOver = false;
while (!gameOver)
{
neuralNetwork.Compute();
onSimulationStep.Invoke();
step++;
if (deltaTime > 0)
{
yield return(new WaitForSeconds(deltaTime));
}
}
yield return(null);
}
public void init(NeuralNetwork.Network nn, int scannerRC)
{
ScanRc = scannerRC;
In_Neurons = In_Synapsis = 0;
int sn0c = nn.Synapsis[StartLayer - 1].Length, sn0i = sn0c;
forEach_Input((int startOff, int cnt) => { In_Neurons += cnt; });
sn0c += (In_Synapsis = In_Neurons * Layers[0].Cnt);
var synA = extendArray(nn, StartLayer - 1, sn0c);
forEach_Input((int startOff, int cnt) => {
for (int i = cnt; i-- > 0;)
{
for (int j = Layers[0].Cnt; j-- > 0;)
{
synA[sn0i++] = nn.setSynapsis(0, startOff + i, StartLayer, Layers[0].Ni1 + j);
}
}
});
int snOc = nn.Synapsis[nn.Synapsis.Length - 1].Length, snOi = snOc;
Out_Neurons = Out_Synapsis = 0;
forEach_Output((int startOff, int cnt) => { Out_Neurons += cnt; });
snOc += (Out_Synapsis = Out_Neurons * Layers[Layers.Count - 1].Cnt);
var synO = extendArray(nn, nn.Synapsis.Length - 1, snOc);
forEach_Output((int startOff, int cnt) => {
for (int i = cnt; i-- > 0;)
{
for (int j = Layers[Layers.Count - 1].Cnt; j-- > 0;)
{
synO[snOi++] = nn.setSynapsis(StartLayer + Layers.Count - 1, Layers[Layers.Count - 1].Ni1 + j, nn.Synapsis.Length, startOff + i);
}
}
});
}
/// <summary>
/// Creates copy of neural network and curves in new instance
/// </summary>
/// <param name="network"></param>
public MyNeuralNetwork(MyNeuralNetwork network)
{
m_network = new Network(network.m_network);
m_curves = new Dictionary<string, ClassicCurve>();
//necessary to do proper copy, otherwise problems with NnIndex (stayed modyfied when Form_modify canceled)
foreach (ClassicCurve curve in network.Curves.Values)
m_curves.Add(curve.ID, new ClassicCurve(curve));
if (network.m_curveSets != null)
{
m_curveSets = new Dictionary<string, MyCurve>();
foreach (MyCurve curve in network.m_curveSets.Values)
m_curveSets.Add(curve.ID, new MyCurve(curve));
//m_curveSets = new List<MyCurve>(network.m_curveSets);
}
}
public NetworkVisualizer(Network network, double[,,] input)
{
this.network = network;
this.input = input;
}
public Benchmark(Network network)
{
this.network = network;
}
public void init(AIConfig[] aic)
{
if (aic == null || aic.Length == 0) //old
{
int[] layers = { Input.Length, 8, 4 };
try {
var r = Random.value; //lazy..
NN = new NeuralNetwork.Network(layers, true, Random.seed);
if (loadWeight)
{
NN.Weights = LoadWeights();
}
} catch (System.Exception e) {
Debug.LogError("NN err: " + e.Message);
}
}
else
{
int mxLayer = -1;
foreach (var c in aic)
{
mxLayer = Mathf.Max(mxLayer, c.Layers.Count + c.StartLayer);
}
int[] layers = new int[mxLayer + 1];
layers[0] = Input.Length;
layers[mxLayer] = 4;
foreach (var c in aic)
{
if (c.enabled)
{
for (int i = c.Layers.Count; i-- > 0;)
{
c.Layers[i].Ni1 = layers[c.StartLayer + i];
layers[c.StartLayer + i] += c.Layers[i].Cnt;
}
}
}
Debug.Log("lc " + layers.Length);
foreach (int i in layers)
{
Debug.Log(" " + i);
}
// try {
var r = Random.value; //lazy..
NN = new NeuralNetwork.Network(layers, Random.seed);
foreach (var c in aic)
{
if (c.enabled)
{
c.init(NN, Motor.Scanner.Count);
}
}
var output = NN.Compute(Input);
/*if(loadWeight) {
* NN.Weights = LoadWeights();
* } */
//} catch(System.Exception e) {
// Debug.LogError("NN err 1: " + e.Message);
// }
}
}
public void init(NeuralNetwork.Network nn)
{
//NN = nn;
NN = new NeuralNetwork.Network(nn);
}
void MutateGene(NeuralNetwork.Network networkToMutate)
{
int geneCount = networkToMutate.GetGeneCount();
for (int i = 0; i < geneToMutate; i++)
{
int geneToTransferId = Random.Range(0, geneCount);
//Look in HiddenLayers
for (int l = 0; l < networkToMutate.hiddenLayers.Length; l++)
{
for (int n = 0; n < networkToMutate.hiddenLayers[l].neurons.Length; n++)
{
int neuronGeneCount = networkToMutate.hiddenLayers[l].neurons[n].GetGeneCount();
//If the gene isn't in this neuron, continue to search
if (geneToTransferId >= neuronGeneCount)
{
geneToTransferId -= neuronGeneCount;
}
//If it's the last one, transfer the activation value
else if (geneToTransferId == neuronGeneCount - 1)
{
networkToMutate.hiddenLayers[l].neurons[n].activationValue += Random.Range(-1, 1);
geneToTransferId = -1;
break;
}
//Else tranfer the weight
else
{
networkToMutate.hiddenLayers[l].neurons[n].weights[geneToTransferId] += Random.Range(-1, 1);
geneToTransferId = -1;
break;
}
}
if (geneToTransferId < 0)
{
break;
}
}
//Look in the output layer
if (geneToTransferId >= 0)
{
for (int n = 0; n < networkToMutate.outputLayer.neurons.Length; n++)
{
int neuronGeneCount = networkToMutate.outputLayer.neurons[n].GetGeneCount();
//If the gene isn't in this neuron, continue to search
if (geneToTransferId >= neuronGeneCount)
{
geneToTransferId -= neuronGeneCount;
}
//If it's the last one, transfer the activation value
else if (geneToTransferId == neuronGeneCount - 1)
{
networkToMutate.outputLayer.neurons[n].activationValue += Random.Range(-1, 1);
geneToTransferId = -1;
break;
}
//Else tranfer the weight
else
{
networkToMutate.outputLayer.neurons[n].weights[geneToTransferId] += Random.Range(-1, 1);
geneToTransferId = -1;
break;
}
}
}
}
}
protected override void Update(GameTime gameTime)
{
this.TargetElapsedTime = TimeSpan.FromMilliseconds(2f);
// Allows the game to exit
if (GamePad.GetState(PlayerIndex.One).Buttons.Back == ButtonState.Pressed)
this.Exit();
if (Agents.Count < 20)
{
if (BestBrain != null&&secondBestBrain!=null)
{
Agent a = new Agent(new Random(), this);
string rep = GenomeEncoder.Encode(BestBrain);
//Console.WriteLine(rep.Length / 32);
string rep2 = GenomeEncoder.Encode(secondBestBrain);
string dr = GenomeEncoder.Combine(rep, rep2);
//Console.WriteLine(dr);
a.brain = GenomeEncoder.Decode(dr);
//Console.WriteLine(dr.Length - rep.Length);
Agents.Add(a);
}
else if (!firstSpawned)
{
Agents.Add(new Agent(new Random(), this));
}
}
else
{
if (!firstSpawned)
{
firstSpawned = true;
}
}
if (Keyboard.GetState(PlayerIndex.One).IsKeyDown(Keys.B))
{
if (!justpressedA){
isPaused = !isPaused;
if (isPaused)
{
Console.WriteLine("Current Best Fitness: " + BestFit);
XmlSerializer ser = new XmlSerializer(typeof(Network));
Console.WriteLine(BestBrain.isTraining);
TextWriter writer = new StreamWriter(@"Ihateyou.txt");
ser.Serialize(writer, BestBrain);
string all = GenomeEncoder.Encode(BestBrain);
//Console.WriteLine(.Length/32);
Network x = GenomeEncoder.Decode(all);
//if(x.HiddenLayer1[2].TargetSynapses[0].Weight==BestBrain.HiddenLayer1[2].TargetSynapses[0].Weight&& x.Inputs[2].TargetSynapses[0].Weight == BestBrain.Inputs[2].TargetSynapses[0].Weight && x.HiddenLayer2[2].RecursiveSynapses[0].Weight == BestBrain.HiddenLayer2[2].RecursiveSynapses[0].Weight)
// {
// Console.WriteLine("I guess they are equal I guess");
//}
// byte[] r = BitConverter.GetBytes(0.34f);
// BitArray a = new BitArray(r);
//Console.WriteLine(BitConverter.ToSingle(GenomeEncoder.ConvertToByte(a),0));
}
justpressedA = true;
}
}
else
{
justpressedA = false;
}
if (Keyboard.GetState(PlayerIndex.One).IsKeyDown(Keys.A))
{
steps = 1;
Console.WriteLine("B Pressed");
}
if (!isPaused) {
for (int x = 0; x < steps;x++){
for (int i = 0; i < Agents.Count; i++)
{
var n = Agents[i];
if (n.Health <= 0)
{
//if (n.Dead != true)
// {
// Agents.Add(new Agent(new Random(), this));
// n.Dead = true;
// }
//Console.WriteLine(i + "'s fitness was " + n.Fitness);
if (n.Fitness > BestFit) {
secondBestBrain = BestBrain;
SecondBestFit = BestFit;
BestFit = n.Fitness;
BestBrain = n.brain;
}
else if(n.Fitness>SecondBestFit)
{
secondBestBrain = n.brain;
SecondBestFit = n.Fitness;
//Console.WriteLine("Found second: " + n.Fitness);
}
Agents.Remove(n);
continue;
}
n.Tick();
}
}
}
// TODO: Add your update logic here
//Angle = (Angle + 1)%360;
base.Update(gameTime);
}
//todo - class-ify optimiser so it can haz parameters and not look shit -- also, ya know, actually do it properly
void geneticOptimisation(Scorer t1, Scorer t2)
{
NeuralNetwork.Network n1 = t1.Ctrl.NN, n2 = t2.Ctrl.NN;
if (n1 != n2)
{
n2.copyTo(n1);
}
/*
* for(int layer = n1.Neurons.Length; --layer > 0; ) {
* int nl = n1.Neurons[layer].Length;
* for(int ri = Mathf.CeilToInt((float)nl * Random.Range(0.7f, 0.9f) ); ri-- > 0; ) {
* int ni = Random.Range(0, nl); //chance of duplicate - i suspect tracking this would be more cost than worth
* n1.Neurons[layer][ni].bias = n2.Neurons[layer][ni].bias;
* }
* var s1 = n1.Synapsis[layer - 1];
* var s2 = n2.Synapsis[layer - 1];
* int sl = s1.Length;
* for(int ri = Mathf.CeilToInt((float)s1.Length * Random.Range(0.7f, 0.9f) ); ri-- > 0; ) {
* int si = Random.Range(0, s1.Length); //chance of duplicate - i suspect tracking this would be more cost than worth
* s1[si].weight = s2[si].weight;
* }
* } */
int nl = n1.Neurons.Length;
for (int ri = Mathf.CeilToInt((float)nl * Random.Range(0.0f, 0.1f)); ri-- > 0;) //todo -- this now affecting input biases (does nothing)
{
int ni = Random.Range(0, nl); //chance of duplicate - i suspect tracking this would be more cost than worth
var m = Random.Range(-1.0f, 1.0f);
if (Random.value > 0.9f)
{
n1.Neurons[ni].bias = m;
}
else
{
m *= Mathf.Abs(m);
n1.Neurons[ni].bias = n1.Neurons[ni].bias + m * 0.05f; //NeuralNetwork.ActivationMethods.HyperbolidTangent(
}
}
for (int layer = n1.Synapsis.Length; layer-- > 0;)
{
var s1 = n1.Synapsis[layer];
int sl = s1.Length;
for (int ri = Mathf.CeilToInt((float)s1.Length * Random.Range(0.0f, 0.1f)); ri-- > 0;)
{
int si = Random.Range(0, s1.Length); //chance of duplicate - i suspect tracking this would be more cost than worth
var m = Random.Range(-1.0f, 1.0f);
if (Random.value > 0.9f)
{
s1[si].weight = m;
}
else
{
m *= Mathf.Abs(m);
s1[si].weight = s1[si].weight + m * 0.05f; // NeuralNetwork.ActivationMethods.HyperbolidTangent(
}
}
}
}
public static string Encode(Network input)
{
StringBuilder output = new StringBuilder();
foreach(Neuron n in input.Inputs)
{
foreach(Synapse t in n.TargetSynapses)
{
BitArray x = new BitArray(BitConverter.GetBytes(t.Weight));
foreach(bool i in x)
{
output.Append(i?"1":"0");
}
}
}
foreach (Neuron n in input.HiddenLayer1)
{
foreach (Synapse t in n.TargetSynapses)
{
BitArray u = new BitArray(BitConverter.GetBytes(t.Weight));
foreach (bool i in u)
{
output.Append(i ? "1" : "0");
}
}
foreach (Synapse t in n.RecursiveSynapses)
{
BitArray u = new BitArray(BitConverter.GetBytes(t.Weight));
foreach (bool i in u)
{
output.Append(i ? "1" : "0");
}
}
BitArray x = new BitArray(BitConverter.GetBytes(n.Threshold));
foreach (bool i in x)
{
output.Append(i ? "1" : "0");
}
}
foreach (Neuron n in input.HiddenLayer2)
{
foreach (Synapse t in n.TargetSynapses)
{
BitArray g = new BitArray(BitConverter.GetBytes(t.Weight));
foreach (bool i in g)
{
output.Append(i ? "1" : "0");
}
}
foreach (Synapse t in n.RecursiveSynapses)
{
BitArray x = new BitArray(BitConverter.GetBytes(t.Weight));
foreach (bool i in x)
{
output.Append(i ? "1" : "0");
}
}
BitArray u = new BitArray(BitConverter.GetBytes(n.Threshold));
foreach (bool i in u)
{
output.Append(i ? "1" : "0");
}
}
foreach (Neuron n in input.HiddenLayer3)
{
foreach (Synapse t in n.TargetSynapses)
{
BitArray u= new BitArray(BitConverter.GetBytes(t.Weight));
foreach (bool i in u)
{
output.Append(i ? "1" : "0");
}
//Console.WriteLine(t.Weight);
}
foreach (Synapse t in n.RecursiveSynapses)
{
BitArray u = new BitArray(BitConverter.GetBytes(t.Weight));
foreach (bool i in u)
{
output.Append(i ? "1" : "0");
}
}
BitArray x = new BitArray(BitConverter.GetBytes(n.Threshold));
foreach (bool i in x)
{
output.Append(i ? "1" : "0");
}
//Console.WriteLine(n.Threshold);
}
foreach (Neuron n in input.Outputs)
{
BitArray u = new BitArray(BitConverter.GetBytes(n.Threshold));
foreach (bool i in u)
{
output.Append(i ? "1" : "0");
}
//Console.WriteLine(n.Threshold);
}
return output.ToString();
}
void Breed(NeuralNetwork.Network mother, NeuralNetwork.Network father, NeuralNetwork.Network child)
{
//For each hidden layer
for (int l = 0; l < mother.hiddenLayers.Length; l++)
{
//For each neuron in the layer
for (int n = 0; n < mother.hiddenLayers[l].neurons.Length; n++)
{
//For each weights
for (int w = 0; w < mother.hiddenLayers[l].neurons[n].weights.Length; w++)
{
//Take randomly the gene from the father or the mother
if (Random.Range(0.0f, 1.0f) > 0.5f)
{
child.hiddenLayers[l].neurons[n].weights[w] = mother.hiddenLayers[l].neurons[n].weights[w];
}
else
{
child.hiddenLayers[l].neurons[n].weights[w] = father.hiddenLayers[l].neurons[n].weights[w];
}
}
//Same for activation value of the neuron
if (Random.Range(0.0f, 1.0f) > 0.5f)
{
child.hiddenLayers[l].neurons[n].activationValue = mother.hiddenLayers[l].neurons[n].activationValue;
}
else
{
child.hiddenLayers[l].neurons[n].activationValue = father.hiddenLayers[l].neurons[n].activationValue;
}
}
}
///Output Layer
//For each neuron in the layer
for (int n = 0; n < mother.outputLayer.neurons.Length; n++)
{
//For each weights
for (int w = 0; w < mother.outputLayer.neurons[n].weights.Length; w++)
{
//Take randomly the gene from the father or the mother
if (Random.Range(0.0f, 1.0f) > 0.5f)
{
child.outputLayer.neurons[n].weights[w] = mother.outputLayer.neurons[n].weights[w];
}
else
{
child.outputLayer.neurons[n].weights[w] = father.outputLayer.neurons[n].weights[w];
}
}
//Same for activation value of the neuron
if (Random.Range(0.0f, 1.0f) > 0.5f)
{
child.outputLayer.neurons[n].activationValue = mother.outputLayer.neurons[n].activationValue;
}
else
{
child.outputLayer.neurons[n].activationValue = father.outputLayer.neurons[n].activationValue;
}
}
}
public Network(Network network)
{
m_layers = new Layer[network.m_layers.Length];
for (int i = 0; i < m_layers.Length; i++)
m_layers[i] = new Layer(network[i]);
}
/// <summary>
/// Creates default neural network accroding to settings
/// </summary>
private void InitializeNetwork()
{
double[][] thresholds = new double[][] { new double[m_hiddenLayerSize], new double[m_outputSize] };
for (int i = 0; i < thresholds.Length; i++)
for (int j = 0; j < thresholds[i].Length; j++)
thresholds[i][j] = StaticRandom.RandomDouble() * (-1);
m_network = new Network(m_inputSize, new int[] { m_hiddenLayerSize, m_outputSize }, thresholds,
new IActivationFunction[] { new HyperbolicTangentFunction(), new SigmoidFunction() });
}
