public PredictorMLP(int aDimension, int aInputLength)
{
dimension = aDimension;
inputLength = aInputLength;
mlp = new MLPDll(new int[] { 30, dimension }, dimension * inputLength);
r = new Random();
mu = 0.0001;
trainCount = 0;
}
public GridNeuralController(IGridModelSimulator model, IObstaclePositionProvider obstacle, ICarPositionProvider start, IFinishPositionProvider finish)
{
this.model = model;
//controller = new MLPDll(new int[] { 40, 1 }, 5, true);//4 bemenet a state, 1 kimenet az input
controllerOriginal = new MLPDll(new int[] { 40, 1 }, 5, false);//4 bemenet a state, 1 kimenet az input
obstacleProvider = obstacle;
carStateProvider = start;
finishStateProvider = finish;
trainingStopped = true;
mu = 0.005;
}
public NeuralController(IModelSimulator model, IObstaclePositionProvider obstacle, ICarPositionProvider start, IFinishPositionProvider finish)
{
this.model = model;
if (INPUT_TYPE == inputType.wheelAngle)
{
controller = new MLPDll(new int[] { 20, 1 }, 4);//4 bemenet a state, 1 kimenet az input
}
else if (INPUT_TYPE == inputType.wheelSpeed)
{
controller = new MLPDll(new int[] { 75, 2 }, 4);//4 bemenet a state, 2 kimenet az input
}
obstacleProvider = obstacle;
carStateProvider = start;
finishStateProvider = finish;
trainingStopped = true;
mu = 0.005;
}
public static void SimulateOneStep(MLPDll controller, IModelSimulator model, CarModelState state, out CarModelInput outInput, out CarModelState outState)
{
double[] inputs = new double[4];
inputs[0] = ComMath.Normal(state.Position.X, CarModelState.MIN_POS_X, CarModelState.MAX_POS_X, POSITION_SCALE * MIN_NEURON_VALUE, POSITION_SCALE * MAX_NEURON_VALUE);
inputs[1] = ComMath.Normal(state.Position.Y, CarModelState.MIN_POS_Y, CarModelState.MAX_POS_Y, POSITION_SCALE * MIN_NEURON_VALUE, POSITION_SCALE * MAX_NEURON_VALUE);
inputs[2] = ComMath.Normal(state.Orientation.X, CarModelState.MIN_OR_XY, CarModelState.MAX_OR_XY, MIN_NEURON_VALUE, MAX_NEURON_VALUE);
inputs[3] = ComMath.Normal(state.Orientation.Y, CarModelState.MIN_OR_XY, CarModelState.MAX_OR_XY, MIN_NEURON_VALUE, MAX_NEURON_VALUE);
double[] controllerOutputs = controller.Output(inputs);
if (INPUT_TYPE == inputType.wheelAngle)
{
outInput = new CarModelInput();
outInput.Angle = ComMath.Normal(controllerOutputs[0], MIN_NEURON_VALUE, MAX_NEURON_VALUE, CarModelInput.MIN_ANGLE, CarModelInput.MAX_ANGLE);
}
else if (INPUT_TYPE == inputType.wheelSpeed)
{
outInput = new CarModelInput(ComMath.Normal(controllerOutputs[0], MIN_NEURON_VALUE, MAX_NEURON_VALUE, CarModelInput.MIN_SPEED, CarModelInput.MAX_SPEED),
ComMath.Normal(controllerOutputs[1], MIN_NEURON_VALUE, MAX_NEURON_VALUE, CarModelInput.MIN_SPEED, CarModelInput.MAX_SPEED));
//********
//hatrafele tilos mennie
if (outInput.LeftSpeed < 0) outInput.LeftSpeed = 0;
if (outInput.RightSpeed < 0) outInput.RightSpeed = 0;
//********
}
model.SimulateModel(state, outInput, out outState);
}
private double TrainOneEpoch(double mu, out double SumSimCount, out List<CarModelState> innerStates)
{
int maxSimCount = 100;
double sumSimCount = 0;
double error = 0;
innerStates = new List<CarModelState>();
List<double> deltaws = new List<double>();
MLPDll[] controllers = new MLPDll[maxSimCount];
IModelSimulator[] models = new IModelSimulator[maxSimCount];
CarModelState state = carStateProvider.GetCarState();
CarModelInput input = new CarModelInput();
//kimenet kiszamitasa
int simCount = 0;
List<double[]> singleErrors = new List<double[]>();
List<double[]> regularizationErrors = new List<double[]>();
CarModelState laststate;
bool earlyStop;
do
{
controllers[simCount] = new MLPDll(controller);//lemasoljuk
models[simCount] = model.Clone();//a modellt is
laststate = state;
NeuralController.SimulateOneStep(controllers[simCount], models[simCount], state, out input, out state);//vegigszimulaljuk a simCount darab controlleren es modellen
innerStates.Add(state);
//kozbulso hibak kiszamitasa, itt csak az akadalyoktol valo tavolsag "hibajat" vesszuk figyelembe, irany nem szamit -> hibaja 0
regularizationErrors.Add(obstacleFieldErrorGradient(state, simCount));
//minden pont celtol vett tavolsaga
double[] desiredOutput = (double[])finishStateProvider.GetFinishState(simCount);
singleErrors.Add(new double[] { 1*ComMath.Normal(desiredOutput[0] - state.Position.X,CarModelState.MIN_POS_X, CarModelState.MAX_POS_X, MIN_NEURON_VALUE, MAX_NEURON_VALUE),
1*ComMath.Normal(desiredOutput[1] - state.Position.Y,CarModelState.MIN_POS_Y, CarModelState.MAX_POS_Y, MIN_NEURON_VALUE, MAX_NEURON_VALUE),
0.1*ComMath.Normal(desiredOutput[2] - state.Orientation.X,CarModelState.MIN_OR_XY, CarModelState.MAX_OR_XY, MIN_NEURON_VALUE, MAX_NEURON_VALUE),
0.1*ComMath.Normal(desiredOutput[3] - state.Orientation.Y,CarModelState.MIN_OR_XY, CarModelState.MAX_OR_XY, MIN_NEURON_VALUE, MAX_NEURON_VALUE) }
);
++simCount;
earlyStop = false;
if (simCount > 3)
{
double[] err1 = singleErrors[simCount-1];
double[] err2 = singleErrors[simCount-2];
double[] err3 = singleErrors[simCount-3];
double error1, error2, error3;
error1 = error2 = error3 = 0;
for (int i = 0; i < err1.Length; i++)
{
error1 += err1[i] * err1[i];
error2 += err2[i] * err2[i];
error3 += err3[i] * err3[i];
}
earlyStop = ((error1 > error2) && (error3 > error2));
if (earlyStop)
{
//utolso elemet toroljuk
singleErrors.RemoveAt(singleErrors.Count - 1);
regularizationErrors.RemoveAt(regularizationErrors.Count - 1);
innerStates.RemoveAt(innerStates.Count - 1);
--simCount;
}
}
}
while ((simCount < maxSimCount) && !earlyStop);
double[] errors = singleErrors[singleErrors.Count-1];
sumSimCount += simCount;
//hibavisszaterjesztes
for (int i = simCount - 1; i >= 0; --i)
{
double[] sensitibility;
models[i].CalcErrorSensibility(errors, out sensitibility);
double[] inputSensitibility;
if (INPUT_TYPE == inputType.wheelAngle)
{
inputSensitibility = new double[1];
inputSensitibility[0] = sensitibility[6];
}
else if (INPUT_TYPE == inputType.wheelSpeed)
{
inputSensitibility = new double[2];
inputSensitibility[0] = sensitibility[4];
inputSensitibility[1] = sensitibility[5];
}
double[] sensitibility2;
controllers[i].SetOutputError(inputSensitibility);
controllers[i].Backpropagate();
controllers[i].CalculateDeltaWeights();
sensitibility2 = controllers[i].SensitibilityD();
errors[0] = (sensitibility[0] + sensitibility2[0]);
errors[1] = (sensitibility[1] + sensitibility2[1]);
errors[2] = (sensitibility[2] + sensitibility2[2]);
errors[3] = (sensitibility[3] + sensitibility2[3]);
//regularizaciobol szarmazo hiba hozzaadasa
errors[0] += regularizationErrors[i][0];
errors[1] += regularizationErrors[i][1];
}
controller.ClearDeltaWeights();
//sulymodositasok osszegzese
for (int i2 = 0; i2 < simCount; ++i2)
{
controller.AddDeltaWeights(controllers[i2]);
}
float maxdw = controller.MaxDeltaWeight();
//if (maxdw < 50) maxdw = 50;
controller.ChangeWeights(mu / maxdw);
////sulymodositasok osszegzese
//for (int i2 = 0; i2 < simCount; ++i2) //simCount
//{
// int count = 0;
// for (int i = 1; i < controllers[i2]; ++i)
// {
// foreach (INeuron n in controllers[i2].mlp[i])
// {
// foreach (NeuronInput ni in ((Neuron)n).inputs)
// {
// if (deltaws.Count <= count) deltaws.Add(ni.deltaw);
// else deltaws[count] += ni.deltaw;
// ++count;
// }
// }
// }
//}
////legnagyobb sulymodositas ertekenek meghatarozasa, majd ezzel normalas
//double maxdw = 1;
//foreach (double dw in deltaws)
//{
// if (Math.Abs(dw) > maxdw) maxdw = Math.Abs(dw);
//}
//if (maxdw < 50) maxdw = 50;
////sulymodositasok ervenyre juttatasa a controllerben
//int count2 = 0;
//for (int i = 1; i < controller.mlp.Count; ++i)
//{
// foreach (INeuron n in controller.mlp[i])
// {
// foreach (NeuronInput ni in ((Neuron)n).inputs)
// {
// ni.w += mu * deltaws[count2] / maxdw;
// ++count2;
// }
// }
//}
SumSimCount = sumSimCount;
return error;
}
public NeuralModelSimulator()
{
mlp = new MLPDll(new int[] { 50, 4 }, 7);
}
public static void SimulateOneStep(MLPDll controller, IGridModelSimulator model, GridCarModelState state, out GridCarModelInput outInput, out GridCarModelState outState)
{
double[] inputs = new double[5];
inputs[0] = ComMath.Normal(state.TargetDist, GridCarModelState.MIN_DIST, GridCarModelState.MAX_DIST, 0, MAX_NEURON_VALUE);
inputs[1] = ComMath.Normal(state.TargetOrientation.X, GridCarModelState.MIN_OR_XY, GridCarModelState.MAX_OR_XY, MIN_NEURON_VALUE, MAX_NEURON_VALUE);
inputs[2] = ComMath.Normal(state.TargetOrientation.Y, GridCarModelState.MIN_OR_XY, GridCarModelState.MAX_OR_XY, MIN_NEURON_VALUE, MAX_NEURON_VALUE);
inputs[3] = ComMath.Normal(state.TargetFinishOrientation.X, GridCarModelState.MIN_OR_XY, GridCarModelState.MAX_OR_XY, MIN_NEURON_VALUE, MAX_NEURON_VALUE);
inputs[4] = ComMath.Normal(state.TargetFinishOrientation.Y, GridCarModelState.MIN_OR_XY, GridCarModelState.MAX_OR_XY, MIN_NEURON_VALUE, MAX_NEURON_VALUE);
double[] controllerOutputs = controller.Output(inputs);
outInput = new GridCarModelInput();
outInput.Angle = ComMath.Normal(controllerOutputs[0], MIN_NEURON_VALUE, MAX_NEURON_VALUE, GridCarModelInput.MIN_ANGLE, GridCarModelInput.MAX_ANGLE);
model.SimulateModel(state, outInput, out outState);
}
public void AddDeltaWeights(MLPDll src1)
{
AddDeltaWeights(src1.mlp, this.mlp, this.mlp);
}
private static double TrainOneEpoch(MLPDll controller, IGridModelSimulator model, ICarPositionProvider cps, IFinishPositionProvider fps, IObstaclePositionProvider ops, double mu, int maxSimCount, out double SumSimCount, out List<GridCarModelState> innerStates)
{
double sumSimCount = 0;
double error = 0;
innerStates = new List<GridCarModelState>();
List<double> deltaws = new List<double>();
MLPDll[] controllers = new MLPDll[maxSimCount];
IGridModelSimulator[] models = new IGridModelSimulator[maxSimCount];
GridCarModelState state = GridCarModelState.FromCarModelState(cps.GetCarState());
GridCarModelInput input = new GridCarModelInput();
//kimenet kiszamitasa
int simCount = 0;
List<double[]> singleErrors = new List<double[]>();
List<double[]> regularizationErrors = new List<double[]>();
GridCarModelState laststate;
bool earlyStop;
do
{
if (simCount == 0) controllers[simCount] = new MLPDll(controller);//lemasoljuk
else controllers[simCount] = new MLPDll(controllers[simCount - 1]);
models[simCount] = model.Clone();//a modellt is
laststate = state;
GridNeuralController.SimulateOneStep(controllers[simCount], models[simCount], state, out input, out state);//vegigszimulaljuk a simCount darab controlleren es modellen
innerStates.Add(state);
//kozbulso hibak kiszamitasa, itt csak az akadalyoktol valo tavolsag "hibajat" vesszuk figyelembe, irany nem szamit -> hibaja 0
regularizationErrors.Add(obstacleFieldErrorGradient(ops, state, simCount));
//minden pont celtol vett tavolsaga
double disterror = ComMath.Normal(state.TargetDist, GridCarModelState.MIN_DIST, GridCarModelState.MAX_DIST, 0, 1);
double orientationerror = disterror;
if (orientationerror < 0.2) orientationerror = 0;
double finishorientationerror = disterror;
if (finishorientationerror > 0.05) finishorientationerror = 0;
else finishorientationerror = 1;
double finishX = Math.Cos(Math.PI - fps.GetFinishState(simCount).Angle);
double finishY = Math.Sin(Math.PI - fps.GetFinishState(simCount).Angle);
singleErrors.Add(new double[] { -disterror * MAX_NEURON_VALUE ,
orientationerror*ComMath.Normal(1 - state.TargetOrientation.X,GridCarModelState.MIN_OR_XY, GridCarModelState.MAX_OR_XY, MIN_NEURON_VALUE, MAX_NEURON_VALUE),
orientationerror*ComMath.Normal(0 - state.TargetOrientation.Y,GridCarModelState.MIN_OR_XY, GridCarModelState.MAX_OR_XY, MIN_NEURON_VALUE, MAX_NEURON_VALUE),
finishorientationerror*ComMath.Normal(finishX - state.TargetFinishOrientation.X,GridCarModelState.MIN_OR_XY, GridCarModelState.MAX_OR_XY, MIN_NEURON_VALUE, MAX_NEURON_VALUE),
finishorientationerror*ComMath.Normal(finishY - state.TargetFinishOrientation.Y,GridCarModelState.MIN_OR_XY, GridCarModelState.MAX_OR_XY, MIN_NEURON_VALUE, MAX_NEURON_VALUE) }
);
++simCount;
earlyStop = false;
if (simCount > 3)
{
double[] err1 = singleErrors[simCount - 1];
double[] err2 = singleErrors[simCount - 2];
double[] err3 = singleErrors[simCount - 3];
double error1, error2, error3;
error1 = error2 = error3 = 0;
for (int i = 0; i < 1; i++)//err1.Length
{
error1 += err1[i] * err1[i];
error2 += err2[i] * err2[i];
error3 += err3[i] * err3[i];
}
earlyStop = ((error1 > error2) && (error3 > error2));
if (earlyStop)
{
//utolso elemet toroljuk
singleErrors.RemoveAt(singleErrors.Count - 1);
regularizationErrors.RemoveAt(regularizationErrors.Count - 1);
innerStates.RemoveAt(innerStates.Count - 1);
--simCount;
}
}
}
while ((simCount < maxSimCount) && !earlyStop);
double[] errors = singleErrors[singleErrors.Count - 1];
sumSimCount += simCount;
//hibavisszaterjesztes
for (int i = simCount - 1; i >= 0; --i)
{
double[] sensitibility;
models[i].CalcErrorSensibility(errors, out sensitibility);
double[] inputSensitibility;
inputSensitibility = new double[1];
inputSensitibility[0] = sensitibility[5];
double[] sensitibility2;
controllers[i].SetOutputError(inputSensitibility);
controllers[i].Backpropagate();
controllers[i].CalculateDeltaWeights();
sensitibility2 = controllers[i].SensitibilityD();
errors[0] = (sensitibility[0] + sensitibility2[0] + 0.1 * singleErrors[i][0]);
errors[1] = (sensitibility[1] + sensitibility2[1] + 0 * singleErrors[i][1]);
errors[2] = (sensitibility[2] + sensitibility2[2] + 0 * singleErrors[i][2]);
errors[3] = (sensitibility[3] + sensitibility2[3] + singleErrors[i][3]);
errors[4] = (sensitibility[4] + sensitibility2[4] + singleErrors[i][4]);
//regularizaciobol szarmazo hiba hozzaadasa
errors[0] += regularizationErrors[i][0];
errors[1] += regularizationErrors[i][1];
errors[2] += regularizationErrors[i][2];
}
controller.ClearDeltaWeights();
//sulymodositasok osszegzese
for (int i2 = 0; i2 < simCount; ++i2)
{
controller.AddDeltaWeights(controllers[i2]);
}
float maxdw = controller.MaxDeltaWeight();
//if (maxdw < 50) maxdw = 50;
controller.ChangeWeights(mu / maxdw);
////sulymodositasok osszegzese
//for (int i2 = 0; i2 < simCount; ++i2) //simCount
//{
// int count = 0;
// for (int i = 1; i < controllers[i2]; ++i)
// {
// foreach (INeuron n in controllers[i2].mlp[i])
// {
// foreach (NeuronInput ni in ((Neuron)n).inputs)
// {
// if (deltaws.Count <= count) deltaws.Add(ni.deltaw);
// else deltaws[count] += ni.deltaw;
// ++count;
// }
// }
// }
//}
////legnagyobb sulymodositas ertekenek meghatarozasa, majd ezzel normalas
//double maxdw = 1;
//foreach (double dw in deltaws)
//{
// if (Math.Abs(dw) > maxdw) maxdw = Math.Abs(dw);
//}
//if (maxdw < 50) maxdw = 50;
////sulymodositasok ervenyre juttatasa a controllerben
//int count2 = 0;
//for (int i = 1; i < controller.mlp.Count; ++i)
//{
// foreach (INeuron n in controller.mlp[i])
// {
// foreach (NeuronInput ni in ((Neuron)n).inputs)
// {
// ni.w += mu * deltaws[count2] / maxdw;
// ++count2;
// }
// }
//}
SumSimCount = sumSimCount;
return error;
}
public MLPDll(MLPDll copy, bool isWeakening)
{
mlp = copyMLPandChangeWeakening(copy.mlp, isWeakening);
}
public MLPDll(MLPDll copy)
{
mlp = copyMLP(copy.mlp);
}
public void AddDeltaWeights(MLPDll src1)
{
AddDeltaWeights(src1.mlp, this.mlp, this.mlp);
}
public MLPDll(MLPDll copy, bool isWeakening)
{
mlp = copyMLPandChangeWeakening(copy.mlp, isWeakening);
}
public MLPDll(MLPDll copy)
{
mlp = copyMLP(copy.mlp);
}