// Update is called once per frame
void Update()
{
if (Input.GetKeyDown("space"))
{
maxBalanceTime = 0;
ResetBall();
}
if (Input.GetKeyDown("j"))
{
Debug.Log(ann.PrintWeights());
}
Debug.DrawRay(transform.position, this.transform.forward * visibleDistance, Color.red);
Debug.DrawRay(transform.position, this.transform.right * visibleDistance, Color.red);
Debug.DrawRay(transform.position, -this.transform.right * visibleDistance, Color.red);
Debug.DrawRay(transform.position, (Quaternion.AngleAxis(-45, Vector3.up) * this.transform.right) * visibleDistance, Color.red);
Debug.DrawRay(transform.position, (Quaternion.AngleAxis(45, Vector3.up) * -this.transform.right) * visibleDistance, Color.red);
weightsString = ann.PrintWeights();
//
}
IEnumerator LoadTrainingSet()
{
string path = Application.dataPath + "/trainingData.txt";
string line;
if (File.Exists(path))
{
int lineCount = File.ReadAllLines(path).Length;
StreamReader tdf = File.OpenText(path);
List <double> calcOutputs = new List <double>();
List <double> inputs = new List <double>();
List <double> outputs = new List <double>();
for (int i = 0; i < epochs; i++)
{
sse = 0;
tdf.BaseStream.Position = 0;
string currentWeights = ann.PrintWeights();
while ((line = tdf.ReadLine()) != null)
{
string[] data = line.Split(',');
float thisError = 0;
if (System.Convert.ToDouble(data[5]) != 0 && System.Convert.ToDouble(data[6]) != 0)
{
inputs.Clear();
outputs.Clear();
inputs.Add(System.Convert.ToDouble(data[0]));
inputs.Add(System.Convert.ToDouble(data[1]));
inputs.Add(System.Convert.ToDouble(data[2]));
inputs.Add(System.Convert.ToDouble(data[3]));
inputs.Add(System.Convert.ToDouble(data[4]));
double o1 = Map(0, 1, -1, 1, System.Convert.ToSingle(data[5]));
outputs.Add(o1);
double o2 = Map(0, 1, -1, 1, System.Convert.ToSingle(data[6]));
outputs.Add(o2);
calcOutputs = ann.Train(inputs, outputs);
thisError = ((Mathf.Pow((float)(outputs[0] - calcOutputs[0]), 2) +
Mathf.Pow((float)(outputs[1] - calcOutputs[1]), 2))) / 2.0f;
}
sse += thisError;
}
trainingProgress = (float)i / (float)epochs;
sse /= lineCount;
if (lastSSE < sse)
{
ann.LoadWeights(currentWeights);
ann.alpha = Mathf.Clamp((float)ann.alpha - 0.001f, 0.01f, 0.9f);
}
else
{
ann.alpha = Mathf.Clamp((float)ann.alpha + 0.001f, 0.01f, 0.9f);
lastSSE = sse;
}
yield return(null);
}
}
trainingDone = true;
}
private IEnumerator LoadTrainingSet()
{
string path = Application.dataPath + "/CarRacing/trainingData.txt";
string line;
if (File.Exists(path))
{
int lineCount = File.ReadAllLines(path).Length;
StreamReader streamReader = File.OpenText(path);
List <double> inputs = new List <double>();
List <double> outputs = new List <double>();
for (int i = 0; i < epochs; i++)
{
sse = 0;
streamReader.BaseStream.Position = 0;
string currentWeights = ann.PrintWeights();
while ((line = streamReader.ReadLine()) != null)
{
string[] data = line.Split(',');
//if nothing to be learned ignore this line
float thisError = 0;
if (System.Convert.ToDouble(data[5]) != 0 && System.Convert.ToDouble(data[6]) != 0)
{
inputs.Clear();
outputs.Clear();
FillInputs(inputs, data);
FillOutputs(outputs, data);
List <double> calcOutputs = ann.Train(inputs, outputs);
thisError = ((Mathf.Pow((float)(outputs[0] - calcOutputs[0]), 2) +
Mathf.Pow((float)(outputs[1] - calcOutputs[1]), 2))) / 2.0f;
}
sse += thisError;
}
trainingProgress = (float)i / (float)epochs;
sse /= lineCount;
CorrectTraining(currentWeights);
yield return(null);
}
}
finishedTraining = true;
SaveWeightsToFile();
}
void SaveWeightsToFile()
{
string path = Application.dataPath + "/BallWeights.txt";
StreamWriter wf = File.CreateText(path);
wf.WriteLine(ann.PrintWeights());
wf.Close();
}
// Use this for initialization
void Start() {
ann = new ANN(aNNBuilder.inputs, aNNBuilder.hidden, aNNBuilder.outputs, aNNBuilder.neuronsPerHidden, aNNBuilder.alpha, aNNBuilder.hiddenFunction, aNNBuilder.outputFunction, aNNBuilder.useWeightsFromFile, aNNBuilder.folder);
ballStartPos = ball.transform.position;
Time.timeScale = timeScale;
if (aNNBuilder.useWeightsFromFile)
ann.LoadWeightsFromFile();
Debug.Log(ann.PrintWeights());
}
// Update is called once per frame
void Update()
{
if (Input.GetKey("space"))
{
Debug.Log(ann.PrintWeights());
}
if (Input.GetKey(KeyCode.Return))
{
ann.LoadWeights();
}
}
public void SaveWeights()
{
ann = bird.GetANN();
System.IO.StreamWriter sw = new System.IO.StreamWriter(@"C:\Users\Ghost\Desktop\weights.txt");
sw.Write(ann.PrintWeights());
sw.Close();
System.IO.StreamWriter sw2 = new System.IO.StreamWriter(@"C:\Users\Ghost\Desktop\bias.txt");
sw2.Write(ann.PrintBias());
sw2.Close();
}
void SaveWeightsToFile()
{
string path = Application.dataPath + "/spyWeights.txt";
//If file path already exists, delete and make a new one to ensure new set of weights are saved
if (File.Exists(path))
{
File.Delete(path);
}
StreamWriter wf = File.CreateText(path);
wf.WriteLine(ann.PrintWeights());
wf.Close();
}
public void TrainNeuralNetworkForWeightScoreUpdateFromDatabase()
{
ann = new ANN(6, 1, 0, 0, 0.8);
List <double> result;
for (int i = 0; i < 1000; i++)
{
for (int j = 0; j < dataNN.Names.Count; j++)
{
double i0 = dataNN.ScorePixelsBalance[j];
double i1 = dataNN.ScoreUnityVisual[j];
double i2 = dataNN.ScoreLawOfLever[j];
double i3 = dataNN.ScoreIsolationBalance[j];
// insert the scoreFromfeatureScore here ... cosi' only 3 score.... not need perceptiron any more.. since
// qui stavo sbagliando dovevo fare solo 4 e poi tre in perceptron...
// devo cambiare anche total score calculation
// parte difficile e fare regression in tensorflowsharp... speriamo bbbenneee..
double i4 = dataNN.ScoreNNFrontTop[j];
double i5 = dataNN.ScoreMobileNet[j];
double output = dataNN.JUDGE[j];
sumSquareError = 0;
result = Train(i0, i1, i2, i3, i4, i5, output);
sumSquareError += Mathf.Pow((float)result[0] - 0, 2);
}
//Debug.Log("SSE: " + sumSquareError);
string weigths = ann.PrintWeights();
//Debug.Log(weigths);
}
}
/// <summary>
/// Loads the training dataset.
/// </summary>
/// <returns> null (Coroutine). </returns>
private IEnumerator LoadTrainingSet()
{
string dataSetFilePath = Application.dataPath + dataSetFolder + dataSetFileName;
// An instance of training (row in the dataSet).
string instance;
if (File.Exists(dataSetFilePath))
{
Debug.Log("Using training data file found at: " + dataSetFilePath);
int instanceCount = File.ReadAllLines(dataSetFilePath).Length;
StreamReader dataSetFile = File.OpenText(dataSetFilePath);
var calculatedOutputs = new List <double>();
var inputs = new List <double>();
var desiredOutputs = new List <double>();
for (var i = 0; i < epochs; i++)
{
sumSquaredError = 0;
// Set file pointer to beginning of file.
dataSetFile.BaseStream.Position = 0;
string currentWeights = ann.PrintWeights();
// Read one instance (line) at a time until the end of the dataSet.
while ((instance = dataSetFile.ReadLine()) != null)
{
// Separate each feature (column) of the current instance (row).
string[] features = instance.Split(',');
// The error we get from a particular instance.
// If nothing to be learned, ignore this line.
float thisError = 0;
// Ignore instances, where no user input was recorded.
// They provide no useful information.
// TODO: Fix floating point number comparison!?
if (System.Convert.ToDouble(features[5]) != 0 &&
System.Convert.ToDouble(features[6]) != 0)
{
inputs.Clear();
desiredOutputs.Clear();
// TODO: Check that training data and inputs are calculated the same way (rounding, normalizing, etc.).
// Assign the first five features (raycast distances) to inputs.
for (int j = 0; j < 5; j++)
{
inputs.Add(System.Convert.ToDouble(features[j]));
}
// Assigns the remaining two features (user input) to outputs.
for (int j = 5; j < 7; j++)
{
double output = Helpers.Map(0, 1, -1, 1, System.Convert.ToSingle(features[j]));
desiredOutputs.Add(output);
}
// Train the Neural Network.
calculatedOutputs = ann.Train(inputs, desiredOutputs);
// Calculate individual squaredErrors.
float output0ErrorSquared = Mathf.Pow((float)(desiredOutputs[0] - calculatedOutputs[0]), 2);
float output1ErrorSquared = Mathf.Pow((float)(desiredOutputs[1] - calculatedOutputs[1]), 2);
// Calculate averaged sum of squared errors.
thisError = (output0ErrorSquared + output1ErrorSquared) / 2f;
}
sumSquaredError += thisError;
}
// Percentage value.
trainingProgress = (float)i / (float)epochs;
// Calculate average sumOfSquaredErrors.
sumSquaredError /= instanceCount;
AdaptLearning(currentWeights);
yield return(null);
}
}
else
{
Debug.LogError("No training data file found at: " + dataSetFilePath);
}
trainingDone = true;
if (!loadWeightsFromFile)
{
SaveWeightsToFile();
}
}
// Method to perform the ANN training using data collected from the player.
IEnumerator LoadTrainingSet()
{
string path = Application.dataPath + "/trainingData.txt";
string line;
if (File.Exists(path))
{
int lineCount = File.ReadAllLines(path).Length;
StreamReader tdf = File.OpenText(path);
List <double> calcOutputs = new List <double>();
List <double> inputs = new List <double>();
List <double> outputs = new List <double>();
//Loop through the epochs
for (int i = 0; i < epochs; i++)
{
//set file pointer to beginning of file
sse = 0;
tdf.BaseStream.Position = 0;
//Get the current weight comma separated string values from the ANN object
string currentWeights = ann.PrintWeights();
// Load the training data, line by line
while ((line = tdf.ReadLine()) != null)
{
string[] data = line.Split(',');
//if nothing to be learned ignore this line
float thisError = 0;
//We are leaving out those data where we have training labels or y (translation and rotation values) with values zero to reduce the data set.
if (System.Convert.ToDouble(data[5]) != 0 && System.Convert.ToDouble(data[6]) != 0) //If translation and rotation outputs in training data are not zero
{
//Clear out lists from previous row
inputs.Clear();
outputs.Clear();
//Add training input data to the inputs to the ANN
inputs.Add(System.Convert.ToDouble(data[0]));
inputs.Add(System.Convert.ToDouble(data[1]));
inputs.Add(System.Convert.ToDouble(data[2]));
inputs.Add(System.Convert.ToDouble(data[3]));
inputs.Add(System.Convert.ToDouble(data[4]));
//Map labels to range (0,1) for efficient training
double o1 = Map(0, 1, -1, 1, System.Convert.ToSingle(data[5]));
outputs.Add(o1);
double o2 = Map(0, 1, -1, 1, System.Convert.ToSingle(data[6]));
outputs.Add(o2);
//Calculated output (y-hat)
calcOutputs = ann.Train(inputs, outputs);
//Sum squared Error value: for both labels
thisError = ((Mathf.Pow((float)(outputs[0] - calcOutputs[0]), 2) +
Mathf.Pow((float)(outputs[1] - calcOutputs[1]), 2))) / 2.0f;
}
//Add this to cumulative SSE for the epoch
sse += thisError;
}
//Percentage training to display on screen
trainingProgress = (float)i / (float)epochs;
// Average SSE
sse /= lineCount;
//If sse isn't better then reload previous set of weights and decrease alpha. This adaptive training to let the ANN move out
// of local optima and hence find global optima.
if (lastSSE < sse)
{
ann.LoadWeights(currentWeights);
ann.alpha = Mathf.Clamp((float)ann.alpha - 0.001f, 0.01f, 0.9f);
}
else //increase alpha
{
ann.alpha = Mathf.Clamp((float)ann.alpha + 0.001f, 0.01f, 0.9f);
lastSSE = sse;
}
yield return(null); //Allow OnGUI some time to update on-screen values
}
}
//Training done, save weights
trainingDone = true;
SaveWeightsToFile();
}
private void QLearning()
{
/// Counting timer
resetTimer--;
qs.Clear();
qs = SoftMax(ann.CalcOutput(states));
float maxQ = qs.Max();
int maxQIndex = qs.ToList().IndexOf(maxQ);
/// Counting exploring output values
exploreRate = Mathf.Clamp(exploreRate - exploreDecay, minExploreRate, maxExploreRate);
if (Random.Range(0, 100) < exploreRate)
{
maxQIndex = Random.Range(0, 4);
}
/// Moving the vehicle using output values
/// Move forward
if (maxQIndex == 0)
{
rb.AddForce(this.transform.forward * Mathf.Clamp(qs[maxQIndex], 0f, 1f) * 300f);
}
if (maxQIndex == 1)
{
rb.AddForce(this.transform.forward * Mathf.Clamp(qs[maxQIndex], 0f, 1f) * -300f);
}
/// Turning
if (maxQIndex == 2)
{
this.transform.Rotate(0, Mathf.Clamp(qs[maxQIndex], 0f, 1f) * 2f, 0, 0);
}
if (maxQIndex == 3)
{
this.transform.Rotate(0, Mathf.Clamp(qs[maxQIndex], 0f, 1f) * -2f, 0, 0);
}
RewardFunction();
/// Setting replay memory
Replay lastMemory = new Replay(states, reward);
Rewards.Add(reward);
if (replayMemory.Count > mCapacity)
{
replayMemory.RemoveAt(0);
}
replayMemory.Add(lastMemory);
/// Training through replay memory
if (collisionFail || resetTimer == 0 || win || backFail)
{
List <float> QOld = new List <float>();
List <float> QNew = new List <float>();
for (int i = replayMemory.Count - 1; i >= 0; i--)
{
List <float> toutputsOld = new List <float>(); // List of actions at time [t] (present)
List <float> toutputsNew = new List <float>(); // List of actions at time [t + 1] (future)
toutputsOld = SoftMax(ann.CalcOutput(replayMemory[i].states)); // Action in time [t] is equal to NN output for [i] step states in replay memory
float maxQOld = toutputsOld.Max(); // maximum Q value at [i] step is equal to maximum Q value in the list of actions in time [t]
int action = toutputsOld.ToList().IndexOf(maxQOld); // number of action (in list of actions at time [t]) with maximum Q value is setted
QOld.Add(toutputsOld[action]);
float feedback;
if (i == replayMemory.Count - 1) // if it's the end of replay memory or if by any reason it's the end of the sequence (in this case
{ // it's collision fail, timer reset and getting into the source of light) then the
feedback = replayMemory[i].reward; // feedback (new reward) is equal to the reward in [i] replay memory, because it's the end of the
} // sequence and there's no event after to count Bellman's equation
else
{
toutputsNew = SoftMax(ann.CalcOutput(replayMemory[i + 1].states)); // otherwise the action at time [t + 1] is equal to NN output for [i + 1] step states
maxQ = toutputsNew.Max(); // in replay memory and then feedback is equal to the Bellman's Equation
feedback = (replayMemory[i].reward +
discount * maxQ);
}
QNew.Add(feedback);
if (save == true)
{
SaveToFile(QOld, QNew, Rewards, "QValues");
}
float thisError = 0f;
currentWeights = ann.PrintWeights();
toutputsOld[action] = feedback; // then the action at time [t] with max Q value (the best action) is setted as counted feedback
List <float> calcOutputs = ann.Train(replayMemory[i].states, toutputsOld); // value and it's used to train NN for [i] state
for (int j = 0; j < calcOutputs.Count; j++)
{
thisError += (Mathf.Pow((toutputsOld[j] - calcOutputs[j]), 2));
}
thisError = thisError / calcOutputs.Count;
sse += thisError;
}
sse /= replayMemory.Count;
if (lastRewardSum < Rewards.Sum())
{
//ann.LoadWeights(currentWeights);
ann.eta = Mathf.Clamp(ann.eta - 0.001f, 0.1f, 0.4f);
}
else
{
ann.eta = Mathf.Clamp(ann.eta + 0.001f, 0.1f, 0.4f);
lastRewardSum = Rewards.Sum();
}
replayMemory.Clear();
ResetVehicle();
Rewards.Clear();
}
}
IEnumerator LoadTrainingSet()
{
string path = ann.GetPath("trainingData");
string line;
if (File.Exists(path))
{
int lineCount = File.ReadAllLines(path).Length;
StreamReader tdf = File.OpenText(path);
List <double> calcOutputs = new List <double>();
List <double> inputs = new List <double>();
List <double> outputs = new List <double>();
for (int i = 0; i < epochs; i++)
{
// set file pointer to beginning of file
sse = 0;
tdf.BaseStream.Position = 0;
string currentWeights = ann.PrintWeights();
while ((line = tdf.ReadLine()) != null)
{
string[] data = line.Split(',');
// if nothing to be learned ignore this line
float thisError = 0;
if (System.Convert.ToDouble(data[5]) != 0 && System.Convert.ToDouble(data[6]) != 0)
{
inputs.Clear();
outputs.Clear();
inputs.Add(System.Convert.ToDouble(data[0]));
inputs.Add(System.Convert.ToDouble(data[1]));
inputs.Add(System.Convert.ToDouble(data[2]));
inputs.Add(System.Convert.ToDouble(data[3]));
inputs.Add(System.Convert.ToDouble(data[4]));
double out1 = Utils.Map(0, 1, -1, 1, System.Convert.ToSingle(data[5]));
double out2 = Utils.Map(0, 1, -1, 1, System.Convert.ToSingle(data[6]));
outputs.Add(out1);
outputs.Add(out2);
calcOutputs = ann.Train(inputs, outputs);
thisError = ((Mathf.Pow((float)(outputs[0] - calcOutputs[0]), 2) +
Mathf.Pow((float)(outputs[1] - calcOutputs[1]), 2))) / 2.0f;
}
sse += thisError;
}
trainingProgress = ((float)i / (float)epochs) * 100;
sse /= lineCount;
if (lastSSE < sse) // if sse isnt better reload old one and decrease alpha
{
ann.LoadWeights(currentWeights);
ann.alpha = Mathf.Clamp((float)ann.alpha - 0.001f, 0.01f, 0.9f);
}
else // increase alpha
{
ann.alpha = Mathf.Clamp((float)ann.alpha + 0.001f, 0.01f, 0.9f);
lastSSE = sse;
}
yield return(null);
}
tdf.Close();
}
trainingDone = true;
ann.SaveWeightsToFile();
}
IEnumerator LoadTrainingSet()
{
string path = Application.dataPath + "/trainingData.txt";
string line;
if (File.Exists(path))
{
int lineCount = File.ReadAllLines(path).Length;
StreamReader tdf = File.OpenText(path);
List <double> calcOutputs = new List <double>();
List <double> inputs = new List <double>();
List <double> outputs = new List <double>();
for (int i = 0; i < epochs; i++)
{
//set file pointer to beginning of file
sse = 0;
tdf.BaseStream.Position = 0;
string currentWeights = ann.PrintWeights();
while ((line = tdf.ReadLine()) != null)
{
string[] data = line.Split(',');
//if nothing to be learned ignore this line
float thisError = 0;
if (Convert.ToDouble(data[5]) != 0 && Convert.ToDouble(data[6]) != 0) //if the data had user input (active keys pressed)
{
inputs.Clear();
outputs.Clear();
inputs.Add(Convert.ToDouble(data[0]));
inputs.Add(Convert.ToDouble(data[1]));
inputs.Add(Convert.ToDouble(data[2]));
inputs.Add(Convert.ToDouble(data[3]));
inputs.Add(Convert.ToDouble(data[4]));
double o1 = Map(0, 1, -1, 1, Convert.ToSingle(data[5]));
outputs.Add(o1);
double o2 = Map(0, 1, -1, 1, Convert.ToSingle(data[6]));
outputs.Add(o2);
calcOutputs = ann.Train(inputs, outputs);
thisError = ((Mathf.Pow((float)(outputs[0] - calcOutputs[0]), 2) + Mathf.Pow((float)(outputs[1] - calcOutputs[1]), 2))) / 2.0f;
}
sse += thisError;
}
trainingProgress = (float)i / (float)epochs;
sse /= lineCount;
//if sse isn't better then reload previous set of weights
//and decrease alpha
if (lastSSE < sse)
{
ann.LoadWeights(currentWeights);
ann.alpha = Mathf.Clamp((float)ann.alpha - 0.001f, 0.01f, 0.9f);
}
else
{
ann.alpha = Mathf.Clamp((float)ann.alpha + 0.001f, 0.01f, 0.9f);
}
yield return(null);
}
}
trainingDone = true;
SaveWeightsToFile();
}