public void CheckAccuracy()
{
Random rand = new Random(31415);
float[] GetSample()
{
var s = new float[3];
s[0] = rand.Next(2);
s[1] = rand.Next(2);
s[2] = ((int)s[0]) ^ ((int)s[1]);
return(s);
}
float Normalize(float value)
{
return(value * 2f - 1);
}
const int inputCount = 2;
const int hiddenCount = 5;
const int outputCount = 1;
var referenceNetwork = new ReferenceFFNetwork(inputCount, hiddenCount, outputCount);
var neatNetwork = new Network(GetNeatNetworkConnections(referenceNetwork.GetWeights(), inputCount, hiddenCount, outputCount),
inputCount + BiasCount,
outputCount,
new DummyNeatChromosomeEncoder());
var sample = GetSample();
neatNetwork.Sensors[0] = Normalize(sample[0]);
neatNetwork.Sensors[1] = Normalize(sample[1]);
neatNetwork.Activate();
var err = MathF.Abs(Normalize(neatNetwork.Effectors[0]) - sample[2]);
for (var i = 0; i < 100; i++)
{
var s = GetSample();
neatNetwork.ActivateAndTrain(s, 0.1f);
referenceNetwork.Train(new[] { s }, 0.1f);
}
neatNetwork.Sensors[0] = Normalize(sample[0]);
neatNetwork.Sensors[1] = Normalize(sample[1]);
neatNetwork.Activate();
var referenceNetworkOutput = referenceNetwork.ComputeOutputs(sample.Take(2).ToArray());
Assert.Equal(referenceNetworkOutput[0], neatNetwork.Effectors[0], 7);
Assert.True(err > MathF.Abs(Normalize(neatNetwork.Effectors[0]) - sample[2]));
}
public double DoEvaluation(Network net, HardMaze env, int runcount = 1000)
{
Session session = new Session(env,
new FitnessHandler(env.compute_fitness),
new NWSELib.EventHandler(doEventHandler),
new InstinctActionHandler(HardMaze.createInstinctAction),
new OptimaGestureHandler(env.GetOptimaGestureHandler),
new TaskBeginHandler(env.TaskBeginHandler));
(List <double> obs, List <double> gesture) = env.reset(net);
net.Reset();
double reward = 0.0;
bool end = false;
for (int time = 0; time < runcount; time++)
{
List <double> inputs = new List <double>(obs);
inputs.AddRange(gesture);
List <double> actions = net.Activate(inputs, time, session, reward);
(obs, gesture, actions, reward, end) = env.action(net, actions);
if (end)
{
break;
}
}
net.Fitness = Session.fitnessHandler == null ? 0 : Session.fitnessHandler(net, session);
net.TaskCompleted = end;
return(net.Fitness);
}
private void 主页ToolStripMenuItem_Click(object sender, EventArgs e)
{
if (network == null || network.IsDisposed)
{
network = new Network();
network.MdiParent = this;
network.FormBorderStyle = FormBorderStyle.None;
network.Dock = DockStyle.Fill;
network.Show();
}
else
{
network.Activate();
}
}
public void CheckFeedforwardBp()
{
const int inputCount = 2;
const int hiddenCount = 2;
const int outputCount = 1;
var referenceNetwork = new ReferenceFFNetwork(inputCount, hiddenCount, outputCount);
var neatNetwork = new Network(GetNeatNetworkConnections(referenceNetwork.GetWeights(), inputCount, hiddenCount, outputCount),
inputCount + BiasCount,
outputCount,
new DummyNeatChromosomeEncoder());
var sampleNetworkOutput = referenceNetwork.ComputeOutputs(new[] { 0f, 1f });
neatNetwork.Sensors[0] = 1f;
neatNetwork.Sensors[1] = 0f;
neatNetwork.Activate();
var neatNetworkOutputBeforeTraining = neatNetwork.Effectors.ToArray();
//the sample network and the NEAT network are identical
Assert.Equal(sampleNetworkOutput, neatNetworkOutputBeforeTraining);
referenceNetwork.Train(new[] { new[] { 0f, 1f, 1f } }, 0.3f);
neatNetwork.ActivateAndTrain(new[] { 1f, 0f, 1f }, 0.3f);
//NEAT network effector values have not changed after training before Activate call
Assert.Equal(neatNetworkOutputBeforeTraining, neatNetwork.Effectors.ToArray());
sampleNetworkOutput = referenceNetwork.ComputeOutputs(new[] { 0f, 1f });
neatNetwork.Activate();
//backpropagation works fine
Assert.Equal(sampleNetworkOutput, neatNetwork.Effectors.ToArray());
Assert.NotEqual(neatNetworkOutputBeforeTraining, neatNetwork.Effectors.ToArray());
}
public void ActivateTest_HiddenNotConnected_Expected_True()
{
// Create network and leave hidden layer not connected
Network net = new Network();
Node inNode1 = new Node(Node.ENodeType.SENSOR, 1);
Node inNode2 = new Node(Node.ENodeType.SENSOR, 2);
Node hiddenNode = new Node(Node.ENodeType.HIDDEN, 3);
Node outNode1 = new Node(Node.ENodeType.OUTPUT, 4);
Node outNode2 = new Node(Node.ENodeType.OUTPUT, 5);
net.AddNodes(new Node[] { inNode1, inNode2, hiddenNode, outNode1, outNode2 });
net.AddLink(new Link(inNode1, outNode1, false, 1.0f));
net.AddLink(new Link(inNode2, outNode2, false, 1.0f));
Assert.AreEqual(true, net.Activate());
}
public void ActivateTest_OneOutputNotConnected_Expected_False()
{
// Create network and leave one of the output nodes unconnected
Network net = new Network();
Node inNode1 = new Node(Node.ENodeType.SENSOR, 1);
Node inNode2 = new Node(Node.ENodeType.SENSOR, 2);
Node hiddenNode = new Node(Node.ENodeType.HIDDEN, 3);
Node outNode1 = new Node(Node.ENodeType.OUTPUT, 4);
Node outNode2 = new Node(Node.ENodeType.OUTPUT, 5);
net.AddNodes(new Node[] { inNode1, inNode2, hiddenNode, outNode1, outNode2 });
net.AddLink(new Link(inNode1, hiddenNode, false, 1.0f));
net.AddLink(new Link(inNode2, hiddenNode, false, 1.0f));
net.AddLink(new Link(hiddenNode, outNode1, false, 1.0f));
Assert.AreEqual(false, net.Activate());
}
public void ActivateTest_AllOutputsConnected_Expected_True()
{
// Create network and make all outputs connected to the network
Network net = new Network();
Node inNode1 = new Node(Node.ENodeType.SENSOR, 1);
Node inNode2 = new Node(Node.ENodeType.SENSOR, 2);
Node hiddenNode = new Node(Node.ENodeType.HIDDEN, 3);
Node outNode1 = new Node(Node.ENodeType.OUTPUT, 4);
Node outNode2 = new Node(Node.ENodeType.OUTPUT, 5);
net.AddNodes(new Node[] { inNode1, inNode2, hiddenNode, outNode1, outNode2 });
net.AddLink(new Link(inNode1, hiddenNode, false, 1.0f));
net.AddLink(new Link(inNode2, hiddenNode, false, 1.0f));
net.AddLink(new Link(hiddenNode, outNode1, false, 1.0f));
net.AddLink(new Link(hiddenNode, outNode2, false, 1.0f));
Assert.AreEqual(true, net.Activate());
}
private static void XorTest(Random r)
{
Genome xorGene = new Genome(r);
xorGene.AddNode(new Node(Node.ENodeType.SENSOR, 1));
xorGene.AddNode(new Node(Node.ENodeType.SENSOR, 2));
xorGene.AddNode(new Node(Node.ENodeType.SENSOR, 3));
xorGene.AddNode(new Node(Node.ENodeType.OUTPUT, 4));
xorGene.AddConnectionGene(1, 4, 0.0f);
xorGene.AddConnectionGene(2, 4, 0.0f);
xorGene.AddConnectionGene(3, 4, 0.0f);
// Run simulation
Simulation sim = new Simulation(r, xorGene, 150);
sim.Parameters.AreConnectionWeightsCapped = false;
sim.Parameters.MaxWeight = 1.0f;
sim.Parameters.WeightMutationPower = 2.5f;
int numberOfRuns = 200;
bool solutionFound = false;
Genome bestGenome = null;
float[] output = { 0.0f, 0.0f, 0.0f, 0.0f };
float[] bestOutput = new float[4];
float[,] input =
{
{ 1.0f, 0.0f, 0.0f },
{ 1.0f, 0.0f, 1.0f },
{ 1.0f, 1.0f, 0.0f },
{ 1.0f, 1.0f, 1.0f }
};
for (int i = 0; i < numberOfRuns; ++i)
{
double epochBestFitness = 0.0f;
float avgConnectionGenes = 0.0f;
// Evaluate all genomes
foreach (Genome gen in sim.Genomes)
{
Network network = gen.GetNetwork();
//network.ActivationFunction = new NeuralEvolution.ActivationFunctions.ReLU();
bool couldActivate = true;
for (int inputIdx = 0; inputIdx < 4; ++inputIdx)
{
float[] inputRow = new float[3];
for (int k = 0; k < 3; ++k)
{
inputRow[k] = input[inputIdx, k];
}
network.SetInput(inputRow);
if (!network.Activate())
{
couldActivate = false;
}
// Relax network
int networkDepth = network.GetMaxDepth();
for (int relax = 0; relax <= networkDepth; ++relax)
{
network.Activate();
}
output[inputIdx] = network.Output[0].Activation;
network.Reset();
}
float errorsum = (float)(Math.Abs(output[0]) + Math.Abs(1.0 - output[1]) + Math.Abs(1.0 - output[2]) + Math.Abs(output[3]));
if (!couldActivate)
{
errorsum = 4;
}
gen.Fitness = Math.Pow((4.0 - errorsum), 2);
gen.Error = errorsum;
if (gen.Fitness > epochBestFitness)
{
bestOutput[0] = output[0]; bestOutput[1] = output[1]; bestOutput[2] = output[2]; bestOutput[3] = output[3];
epochBestFitness = gen.Fitness;
}
avgConnectionGenes += gen.ConnectionGenes.Count;
if ((output[0] < 0.5f) && (output[1] >= 0.5f) && (output[2] >= 0.5f) && (output[3] < 0.5f))
{
bestOutput[0] = output[0]; bestOutput[1] = output[1]; bestOutput[2] = output[2]; bestOutput[3] = output[3];
bestGenome = gen;
solutionFound = true;
break;
}
}
avgConnectionGenes /= sim.Genomes.Count;
Console.WriteLine(String.Format("Epoch {0} | best: {1} | best output: [{2}, {3}, {4}, {5}] | avg genes: {6} | species: {7}", i, epochBestFitness, bestOutput[0], bestOutput[1], bestOutput[2], bestOutput[3], avgConnectionGenes, sim.Species.Count));
// We found a solution so we can exit already
if (solutionFound)
{
break;
}
// Advance to the next step of simulation
sim.Epoch();
}
if (solutionFound)
{
Console.WriteLine(String.Format("Solution found: [{0}, {1}, {2}, {3}]", bestOutput[0], bestOutput[1], bestOutput[2], bestOutput[3]));
Console.WriteLine(String.Format("Solution network nodes count: {0} | connections count: {1}", bestGenome.Nodes.Count, bestGenome.ConnectionGenes.Count));
}
else
{
Console.WriteLine("Solution NOT found!");
}
}
// cart_and_pole() was take directly from the pole simulator written
// by Richard Sutton and Charles Anderson.
private static int Go_cart(Network net, int max_steps, int thresh, Random rnd)
{
float x, /* cart position, meters */
x_dot, /* cart velocity */
theta, /* pole angle, radians */
theta_dot; /* pole angular velocity */
int steps = 0, y;
float[] input = new float[5]; //Input loading array
float out1;
float out2;
float twelve_degrees = 0.2094384f;
/* if (random_start) {
* x = (lrand48()%4800)/1000.0 - 2.4;
* x_dot = (lrand48()%2000)/1000.0 - 1;
* theta = (lrand48()%400)/1000.0 - .2;
* theta_dot = (lrand48()%3000)/1000.0 - 1.5;
* }
* else */
//x = x_dot = theta = theta_dot = 0.0f;
x = (rnd.Next(0, 4800) / 1000.0f - 2.4f);
x_dot = (rnd.Next(0, 2000) / 1000.0f - 1.0f);
theta = (rnd.Next(0, 400) / 1000.0f - 0.2f);
theta_dot = (rnd.Next(0, 3000) / 1000.0f - 1.5f);
/*--- Iterate through the action-learn loop. ---*/
while (steps++ < max_steps)
{
/*-- setup the input layer based on the four iputs --*/
input[0] = 1.0f; //Bias
input[1] = (x + 2.4f) / 4.8f;
input[2] = (x_dot + 0.75f) / 1.5f;
input[3] = (theta + twelve_degrees) / 0.41f;
input[4] = (theta_dot + 1.0f) / 2.0f;
net.SetInput(input);
// Activate the net
// If it loops, exit returning only fitness of 1 step
if (!(net.Activate()))
{
return(1);
}
/*-- Decide which way to push via which output unit is greater --*/
List <Node> outputs = net.Output;
out1 = outputs[0].Activation;
out2 = outputs[1].Activation;
if (out1 > out2)
{
y = 0;
}
else
{
y = 1;
}
/*--- Apply action to the simulated cart-pole ---*/
Cart_pole(y, ref x, ref x_dot, ref theta, ref theta_dot);
/*--- Check for failure. If so, return steps ---*/
if (x < -2.4 || x > 2.4 || theta < -twelve_degrees || theta > twelve_degrees)
{
return(steps);
}
}
return(steps);
}