public override Tensor Compute(Tensor input)
{
var values = (Tensor1D)input;
var results = new Tensor1D(values.Length);
for (var i = 0; i < values.Length; i++)
{
results[i] = 1f / (1 + (float)Math.Exp(-values[i]));
}
return(results);
}
Tensor1D[,] GenerateTrainingData()
{
var trainingdata = new Tensor1D[Record.Count, 2];
for (var i = 0; i < Record.Count; i++)
{
trainingdata[i, 0] = new Tensor1D(new float[] { Record[i].Speed, Record[i].SensorLeft, Record[i].SensorFront, Record[i].SensorRight });
trainingdata[i, 1] = new Tensor1D(new float[] { Record[i].SteeringAngle, Record[i].Tourgue });
}
return(trainingdata);
}
private Tensor1D PropagateGradient(Tensor1D gradient)
{
var propgrad = new float[InputNumber + 1];
for (var i = 0; i < InputNumber + 1; i++)
{
for (var n = 0; n < Neurons; n++)
{
propgrad[i] += Weights[n][i] * gradient[n];
}
}
return(new Tensor1D(propgrad));
}
private Tensor ComputeOutput(Tensor inputs)
{
var input = (Tensor1D)inputs;
input.AddElement(1);
var scalarprodukttensor = new Tensor1D(Neurons);
for (var i = 0; i < Neurons; i++)
{
var product = Weights[i] * input;
var productsum = product.ElementSum();
scalarprodukttensor[i] = productsum;
}
var outputvalues = Activation.Compute(scalarprodukttensor);
return(outputvalues);
}
public void ComputeErrorTensor(Transition tr, Layer Network)
{
var outputblist = ((Tensor1D)Network.Forward(tr.InputB)).Tolist();
var maxrewardindexnextstep = outputblist.IndexOf(outputblist.Max());
var correctreward = tr.Reward + 0.9f * maxrewardindexnextstep;
var errorarray = new float[tr.OutputA.Length];
var errortensor = new Tensor1D(errorarray);
var output = (Tensor1D)Network.Forward(tr.InputA);
var index = tr.Action.SteeringCount * tr.Action.SteeringNumber + tr.Action.TourqueNumber;
errortensor[index] = output[index] - correctreward;
Network.Backward(errortensor);
}
void FixedUpdate()
{
if (_Recording)
{
var currenttime = Time.time;
if (currenttime - _Time > DeltaTime)
{
_Time = currenttime;
Record.Add(GrabDriveSnapShot());
}
}
if (_NetDrive)
{
var drivecontroller = Car.GetComponent <RearWheelDrive>();
var sensors = Car.GetComponentsInChildren <DetectionLaser>();
drivecontroller.InputControl = false;
var snap = GrabDriveSnapShot();
var input = new Tensor1D(new float[] { snap.Speed, snap.SensorLeft, snap.SensorFront, snap.SensorRight });
var result = (Tensor1D)Network.Forward(input);
drivecontroller.SteeringAngle = result[0];
drivecontroller.Tourque = result[1];
}
else if (_NetSelfTrain && Network != null)
{
var drivecontroller = Car.GetComponent <RearWheelDrive>();
var sensors = Car.GetComponentsInChildren <DetectionLaser>();
if (_SelfTraining)
{
return;
}
var snap = GrabDriveSnapShot();
var speed = drivecontroller.Speed;
var speedhigh = speed > 1.5;
//var rr = 20 * Vector3.Distance(LastPos, Car.transform.position) / 2;
var reward = -(snap.SensorFront + snap.SensorLeft + snap.SensorRight) + speed / 10; // + rr;
if (speedhigh || _LastTimeSpeedHigh == 0)
{
_LastTimeSpeedHigh = Time.time;
}
else if (Time.time - _LastTimeSpeedHigh > 3 || drivecontroller.Collided)
{
AddSample(drivecontroller, -5);
_SelfTraining = true;
StartCoroutine("SelfTrainEpoch");
drivecontroller.Tourque = 0;
drivecontroller.SteeringAngle = 0;
return;
}
AddSample(drivecontroller, reward);
}
else
{
var drivecontroller = Car.GetComponent <RearWheelDrive>();
drivecontroller.InputControl = true;
}
}
void AddSample(RearWheelDrive drivecontroller, float reward)
{
drivecontroller.InputControl = false;
var snap = GrabDriveSnapShot();
var input = new Tensor1D(new float[] { snap.Speed, snap.SensorLeft, snap.SensorFront, snap.SensorRight });
var result = (Tensor1D)Network.Forward(input);
var outlist = result.Tolist();
var steeringdim = _Actions.ActionTable.GetLength(0);
var trouquedim = _Actions.ActionTable.GetLength(1);
//a.SteeringNumber* a.SteeringCount + a.TourgueNumber = index
var maxindex = outlist.IndexOf(outlist.Max());
var _rnd = new System.Random();
var randomdouble = _rnd.NextDouble();
if (0.2 >= randomdouble)
{
maxindex = _rnd.Next(0, outlist.Count);
}
var currenttime = Time.time;
if (currenttime - _ActionTime > 0.2)
{
_ActionTime = currenttime;
}
else
{
return;
}
var steeringindex = maxindex / steeringdim;
var touqueindex = maxindex % steeringdim;
var action = _Actions.ActionTable[steeringindex, touqueindex];
drivecontroller.SteeringAngle = action.SteeringAngle;
drivecontroller.Tourque = Math.Max(action.Tourque, 0);
var transition = new Transition();
if (_LastSnap != null)
{
transition.InputA = _LastInput;
transition.InputB = input;
transition.OutputA = _LastOutput;
transition.OutputB = result;
transition.Reward = PreviousReward;
transition.Action = action;
currenttime = Time.time;
if (currenttime - _Time > 0.2)
{
_Time = currenttime;
_Transitions.Add(transition);
}
}
_LastSnap = snap;
_LastInput = input;
_LastOutput = result;
PreviousReward = reward;
LastPos = Car.transform.position;
}
