public void TestBias()
{
double input = 0.3;
double inpToOutWeight = 0.4;
double biasToOutWeight = 0.5;
var sigmoid = new SigmoidFunction();
var nw = new Network(sigmoid, false);
var inputNode = new Input(input);
var bias = new Bias();
var outputNode = new Perceptron(sigmoid, "Output");
Connection.Create(inpToOutWeight, inputNode, outputNode);
Connection.Create(biasToOutWeight, bias, outputNode);
nw.Nodes = new Node[][] {
new Node[] { inputNode },
new Node[] { outputNode }
};
var nwOut = nw.Nodes[1][0].Output;
//Output
// = Sig(inpToOut.Output + biasToOut.Output)
// = Sig((inpToOut.Weight * inpToOut.GetInput()) + (biasToOutWeight * 1))
// = Sig((inpToOut.Weight * inputNode.Output) + (biasToOutWeight))
// = Sig((this.weight * this.input) + (biasToOutWeight))
var expOut = sigmoid.Calculate((inpToOutWeight * input) + biasToOutWeight);
Assert.AreEqual(nwOut, expOut);
}
/// <param name="checkSuccess">Funcion takes 'expected', 'actual' returns bool wether or not actual can be considerd a success</param>
public static ValidationResult Validate(Network network, IEnumerable<InputExpectedResult> inputResults, Func<double[], double[], bool> checkSuccess)
{
ValidationResult totalResult = null;
foreach(var cur in inputResults) {
double[] actual = network.GetInputResult(cur.Input);
double sse = SumSquaredError(cur.Output, actual);
bool success = checkSuccess(cur.Output, actual);
var res = new ValidationResult(sse, success);
totalResult += res;
}
return totalResult;
}
private static void Main(string[] args) {
var data = DataReader.ReadFromFile("data/iris.data")/*.OrderBy(i => random.Next())*/;
foreach(var entry in data) {
//TODO: it is suboptimal to calculate min and max every time
entry.PetalLength = Normalize(entry.PetalLength, data.Min(i => i.PetalLength), data.Max(i => i.PetalLength));
entry.PetalWidth = Normalize(entry.PetalWidth, data.Min(i => i.PetalWidth), data.Max(i => i.PetalWidth));
entry.SepalLength = Normalize(entry.SepalLength, data.Min(i => i.SepalLength), data.Max(i => i.SepalLength));
entry.SepalWidth = Normalize(entry.SepalWidth, data.Min(i => i.SepalWidth), data.Max(i => i.SepalWidth));
}
var network = new Network(new NeuralNet.TransferFunctions.HyperbolicTangentFunction(), true);
network.FillNetwork(4, 3, 6);
var train = data.Take(TrainSetSize).ToArray();
var inputAndExpectedResuls = train.Select(entry => new InputExpectedResult(entry.AsInput, entry.AsOutput));
var validation = data.Skip(TrainSetSize).ToArray();
var bp = new Backpropagate(network, 0.5, 3);
var trainData = inputAndExpectedResuls.ToArray();
int trains = 0;
double score = 0;
var start = Environment.TickCount;
while(score < 90) {
trains++;
bp.Train(trainData.OrderBy(i => random.Next()).ToArray());
var stats = NetworkValidation.Validate(network, inputAndExpectedResuls, IrisEntry.IsOutputSuccess);
score = stats.SuccessPercentage;
Console.WriteLine($"{trains,-4}" + stats.ToString());
}
Console.WriteLine($"\nTime elapsed: {Environment.TickCount - start}Ms");
Console.WriteLine("Done training");
var trainStats = NetworkValidation.Validate(network, train.Select(entry => new InputExpectedResult(entry.AsInput, entry.AsOutput)), IrisEntry.IsOutputSuccess);
var validateStats = NetworkValidation.Validate(network, validation.Select(entry => new InputExpectedResult(entry.AsInput, entry.AsOutput)), IrisEntry.IsOutputSuccess);
Console.WriteLine($"{trainStats.ToString()} TRAIN");
Console.WriteLine($"{validateStats.ToString()} VALIDATE");
Console.WriteLine($"{(trainStats + validateStats).ToString()} TOTAL");
Console.ReadKey();
}
public NeuralNetBot(Checker myColor, Network network, double? lambda=null, LambdaType? lambdaType=null)
: base(myColor, lambda, lambdaType)
{
Network = network;
}
static void Main(string[] args)
{
var nn = new Network(4, 3, 1);
var result = nn.FeedForward(new double[] { 0.4, 0.6, 0.1, 0.2 });
return;
}
public void TestFillNetwork()
{
int inputs = 2;
int layer1 = 3;
int layer2 = 4;
int outputs = 1;
var nw = new Network(new SigmoidFunction(), true);
nw.FillNetwork(inputs, outputs, layer1, layer2);
Assert.IsTrue(nw.Nodes[0].Length == inputs);
Assert.IsTrue(nw.Nodes[1].Length == layer1);
Assert.IsTrue(nw.Nodes[2].Length == layer2);
Assert.IsTrue(nw.Nodes[3].Length == outputs);
Assert.IsTrue(((Perceptron)nw.Nodes[3][0]).GetIncommingConnections().Count == layer2 + 1); // +1 for bias node
Assert.IsTrue(((Perceptron)nw.Nodes[2][3]).GetIncommingConnections().Count == layer1 + 1); // +1 for bias node
}
public void TestTraining()
{
var sigmoid = new SigmoidFunction();
var net = new Network(sigmoid, true);
net.FillNetwork(2, 2, 2);
net.Nodes[0][0].GetOutgoingConnections()[0].Weight = .15;
net.Nodes[0][0].GetOutgoingConnections()[1].Weight = .2;
net.Nodes[0][1].GetOutgoingConnections()[0].Weight = .25;
net.Nodes[0][1].GetOutgoingConnections()[1].Weight = .3;
net.Nodes[1][0].GetOutgoingConnections()[0].Weight = .4;
net.Nodes[1][0].GetOutgoingConnections()[1].Weight = .45;
net.Nodes[1][1].GetOutgoingConnections()[0].Weight = .5;
net.Nodes[1][1].GetOutgoingConnections()[1].Weight = .55;
var expected = new InputExpectedResult(new double[] { .05, .1 }, new double[] { .01, .99 });
var before = NetworkValidation.Validate(net, new InputExpectedResult[] { expected }, (a, b) => true);
var bp = new Backpropagate(net, 0.5);
bp.Train(new InputExpectedResult[] { expected });
var after = NetworkValidation.Validate(net, new InputExpectedResult[] { expected }, (a, b) => true);
Assert.IsTrue(before.AvgSSE > after.AvgSSE);
}
public void TestNetworkOutput()
{
var rand = new Random();
double input = (double)(rand.NextDouble());
double weight = (double)(rand.NextDouble());
var sigmoid = new SigmoidFunction();
var nw = new Network(sigmoid, false);
var inputNode = new Input(input);
var outputNode = new Perceptron(sigmoid, "Output");
Connection.Create(weight, inputNode, outputNode);
nw.Nodes = new Node[][] {
new Node[] { inputNode },
new Node[] { outputNode }
};
var nwOut = nw.CurOutput()[0];
Assert.AreEqual(outputNode.Output, nwOut);
var expOut = sigmoid.Calculate(input * weight);
Assert.AreEqual(nwOut, expOut);
}
void Start()
{
//MIND:
#if MODULUS_MIND
//ASSUMES SINGLE_THREADED. (ELSE PUT A LOCK HERE)
if (minds == null)
{
minds = new NeuralNet.Network[max_mind];
for (int i = 0; i < max_mind; i++)
{
mind = new NeuralNet.Network(brainConfig,
NeuralNet.Misc.sigmoidNF()); //Create a neural-network with input size for the eyes, and 4 outputs.
//if(mind == null)
// Debug.Log("Mind is nulL!");
minds[i] = mind;
}
}
if (mind == null)
{
brainID = cur_mind;
mind = minds[cur_mind];
cur_mind++;
cur_mind = cur_mind % max_mind;
}
//End locked-area.
//simple_train();
#else
if (mind == null)
{
mind = new NeuralNet.Network(brainConfig,
NeuralNet.Misc.sigmoidNF()); //Create a neural-network with input size for the eyes, and 4 outputs.
//simple_train();
}
#endif
myBodyParts = new List <GameObject>();
body = GetComponent <Rigidbody>();
//body.velocity = new Vector3 (1,0,0);
//body.velocity = (new Vector3(Random.Range(-100, 100), 0, Random.Range(-100, 100))).normalized * 10;
#if ADD_EYE
eye = Creature_bits.newEye();
AddBodyPart(eye, (this.transform.forward + new Vector3(0, .4f, 0)));
#endif
#if ADD_MOUTH
GameObject mouth = Instantiate(mouthPart);
AddBodyPart(mouth, this.transform.forward);
#endif
//max_age = 300*12*2; //300 ~= 5secs, *12 make sit about a minute
max_age = 300; //300frames.
max_health = 100;
max_food = 1000;
age = 0;
food = max_food / 2;
health = max_health;
trueColor = GetComponent <Renderer>().material.color;
}
private void InitNetwork(double learnRate, int microBatchsize, int inputHeight, int outputHeight, int[] hiddenHeights)
{
Network = new Network(TransferFunc, true);
Network.FillNetwork(inputHeight, outputHeight, hiddenHeights);
BackpropTrain = new Backpropagate(Network, learnRate, microBatchsize);
}
private void btnStart_Click(object sender, RoutedEventArgs e)
{
try
{
if (Status == TrainStatus.Running)
{
Status = TrainStatus.Paused;
Save();
return;
}
if (Thread != null && Thread.IsAlive)
Thread.Join(); // Wait for last thread to finish working before starting again.
Termination termination;
if (cbTerminationType.Text == "ByValidationSet")
termination = Termination.ByValidationSet(DataParser.ValidationSet, ToInt(tbValidateCycle, x => x > 0, "Validation cycle must be > 0"));
else if (cbTerminationType.Text == "ByIteration")
termination = Termination.ByIteration(ToInt(tbIterations, x => x > 0, "Iterations must be > 0"));
else
throw new Exception("Did not recognize termination type: " + cbTerminationType.Text);
NetworkParameters parameters = new NetworkParameters()
{
InitialWeightInterval = new Tuple<double, double>(ToDouble(tbInitialWeightMin), ToDouble(tbInitialWeightMax)),
LearningRate = ToDouble(tbLearningRate, x => x > 0, "Learning rate must be > 0"),
LearningRateDecay = ToDouble(tbLearningRateDecay),
Momentum = ToDouble(tbMomentum, x => x >= 0 && x < 1, "Momentum must be in [0,1)"),
MomentumDecay = ToDouble(tbMomentumDecay)
};
string name = ToString(tbName, s => s.Trim() != string.Empty, "Network name cannot be empty.");
if (Status == TrainStatus.Create && Directory.Exists(name))
{
if (MessageBox.Show("Neural Net folder '" + name + "' already exists. Delete contents?\r\nIf not, use a unique name.", "Error", MessageBoxButton.YesNo) == MessageBoxResult.Yes)
(new DirectoryInfo(name)).Delete(true);
else
return;
}
if (Status == TrainStatus.Create)
{
Network = new Network(
name,
ToInt(tbInputs, x => x > 0, "Number of input nodes must be > 0"),
tbHiddens.Text.Split(new char[] { ' ', ',', ';' }, StringSplitOptions.RemoveEmptyEntries).Select(x => ToInt(x, tbHiddens, xx => xx > 0, "Number of hidden nodes must be > 0 per layer")).ToList<int>(),
ToInt(tbOutputs, x => x > 0, "Number of output nodes must be > 0"),
termination,
parameters);
}
else
{
Network.Termination.Type = termination.Type;
Network.Termination.ValidationSet = termination.ValidationSet;
Network.Termination.ValidateCycle = termination.ValidateCycle;
Network.Termination.TotalIterations = termination.TotalIterations;
Network.Parameters = parameters;
Network.Name = name;
}
Status = TrainStatus.Running;
Thread = new Thread(new ThreadStart(DoTrain)) { IsBackground = true };
Thread.Start();
}
catch (Exception x)
{
MessageBox.Show(x.Message, "Error");
}
}
private void btnLoad_Click(object sender, RoutedEventArgs e)
{
Save();
Microsoft.Win32.OpenFileDialog open = new Microsoft.Win32.OpenFileDialog();
open.Filter = "Neural Network|*.net";
open.InitialDirectory = AppDomain.CurrentDomain.BaseDirectory;
open.RestoreDirectory = true;
if (open.ShowDialog().Value)
{
try
{
Network = (Network)Serializer.Deserialize(open.FileName);
PopulateControls(Network);
Status = TrainStatus.Paused;
Title = Network.Name;
}
catch { MessageBox.Show("Could not deserialize " + open.FileName.ToString(), "Error"); }
}
}
public void UpdateProgress(Network network)
{
lbError.Content = (!network.TrueError.HasValue ? "0" : String.Format("{0:f5}", network.TrueError.Value).ToString());
lbIteration.Content = network.Termination.CurrentIteration;
lbTimeElapsed.Content = network.TrainTime.Elapsed.ToString(@"d\.hh\:mm\:ss");
for (; CurrentErrorHistoryIndex < network.ErrorHistory.Count; ++CurrentErrorHistoryIndex)
{
ValidationPlot.AppendAsync(this.Dispatcher, network.ErrorHistory[CurrentErrorHistoryIndex]);
}
}
public void PopulateControls(Network network)
{
tbName.Text = network.Name;
tbInputs.Text = network.Inputs.Count.ToString();
tbHiddens.Text = network.HiddensLayout.Aggregate(string.Empty, (total, i) => total += i.ToString() + " ").Trim();
tbOutputs.Text = network.Outputs.Count.ToString();
tbLearningRate.Text = network.Parameters.LearningRate.ToString();
tbLearningRateDecay.Text = network.Parameters.LearningRateDecay.ToString();
tbMomentum.Text = network.Parameters.Momentum.ToString();
tbMomentumDecay.Text = network.Parameters.MomentumDecay.ToString();
tbInitialWeightMin.Text = network.Parameters.InitialWeightInterval.Item1.ToString();
tbInitialWeightMax.Text = network.Parameters.InitialWeightInterval.Item2.ToString();
cbTerminationType.Text = network.Termination.Type.ToString();
tbIterations.Text = network.Termination.TotalIterations.ToString();
tbValidateCycle.Text = network.Termination.ValidateCycle.ToString();
ValidationPlot.Collection.Clear();
TrainingPlot.Collection.Clear();
UpdateProgress(network);
}
/// <summary>
/// Simulate a game until completion.
/// </summary>
/// <param name="board">Starting board that the bots will play on. This need not be empty!</param>
/// <param name="network">Neural network that provides the AI for gameplay.</param>
/// <returns>Trace of game sequence, each board state stored as a Neural Net Example</returns>
public List<Example> Play(Board board, Network network)
{
Bot allen = new NeuralNetBot(Checker.Blue, network); // <-- you know he will win :)
Bot jason = new NeuralNetBot(Checker.Green, network);
List<Example> trace = new List<Example>();
Turns = 0;
Bot current = allen.MyColor == board.NextPlayer ? allen : jason;
while (!board.IsGameOver)
{
int column;
double score;
current.SelectMove(board, out column, out score);
Log(String.Format("{0} picks column {1} (Score: {2:f2})", (current == allen ? "Allen" : "Jason"), column, score));
board.AddChecker(current.MyColor, column);
Example example = Transform.ToNormalizedExample(board, current.MyColor);
example.Predictions.Add(score);
trace.Add(example);
current = (current == allen ? jason : allen);
++Turns;
}
if (Viewer != null)
Viewer.BatchAddCheckers(Checker.Blue, board.MoveHistory,completedBoard:board);
TotalTurns += Turns;
Checker winner;
if (board.TryGetWinner(out winner))
{
//The game is over, there was a winner.
//This means the last element of "trace" represents a won
//board state (i.e. there is a four-in-a-row with color
//'winner').
if (trace.Count > 0) trace[trace.Count - 1].Predictions[0] = Transform.ToValue(GameResult.Win);
if (trace.Count > 1) trace[trace.Count - 2].Predictions[0] = Transform.ToValue(GameResult.Loss);
if (winner == allen.MyColor)
{
Log("WINNER: Allen");
++AllenWon;
}
else
{
Log("WINNER: Jason");
++JasonWon;
}
}
else
{
if (trace.Count > 0) trace[trace.Count - 1].Predictions[0] = Transform.ToValue(GameResult.Draw);
if (trace.Count > 1) trace[trace.Count - 2].Predictions[0] = Transform.ToValue(GameResult.Draw);
Log("TIE");
++Ties;
}
++TotalGames;
Log(string.Format("Turns: {0} ({1:f2})", Turns, (double)TotalTurns / TotalGames));
Log(string.Format("Allen: {0}({1:f2}) Jason: {2}({3:f2}) Ties {4}({5:f2}) TOTAL: {6}", AllenWon, (double)AllenWon / TotalGames, JasonWon, (double)JasonWon / TotalGames, Ties, (double)Ties / TotalGames, TotalGames));
Log("");
List<Example> trace1 = new List<Example>(), trace2 = new List<Example>();
for (int i = 0; i < trace.Count; ++i)
{
if (i % 2 == 0) trace1.Add(trace[i]);
else trace2.Add(trace[i]);
}
double lambda = .7;
double alpha = .1;
double gamma = .5;
UpdateTraceLabels(trace1, lambda, alpha, gamma);
UpdateTraceLabels(trace2, lambda, alpha, gamma);
return trace1.Union(trace2).ToList();
}
public Trainer(NetworkGenerator gui)
{
this.gui = gui;
Network = gui.Network;
}
public Trainer(Network network)
{
Network = network;
}
public void TestGetInputResult()
{
var rand = new Random();
double input = (double)(rand.NextDouble() * 100 + 1);
double weight = (double)(rand.NextDouble() * 2);
var sigmoid = new SigmoidFunction();
var nw = new Network(sigmoid, false);
var inputNode = new Input(input);
var outputNode = new Perceptron(sigmoid, "Output");
Connection.Create(weight, inputNode, outputNode);
nw.Nodes = new Node[][] {
new Node[] { inputNode },
new Node[] { outputNode }
};
var nwOut = nw.GetInputResult(input)[0];
//Output
// = Sig(inpToOut.Output)
// = Sig(inpToOut.Weight * inpToOut.GetInput())
// = Sig(inpToOut.Weight * inputNode.Output)
// = Sig(this.weight * this.input)
var expOut = sigmoid.Calculate(weight * input);
Assert.AreEqual(nwOut, expOut);
}