private static string[] ColumnNames(bool withLabel, ModelValue prediction)
{
var columns = new List <string>();
var data = new ModelDataSet();
for (int i = 0; i < data.Features(); i++)
{
columns.Add(data.Name(i));
}
if (withLabel)
{
switch (prediction)
{
case ModelValue.Action: columns.Add("Action"); break;
case ModelValue.Angle: columns.Add("FaceAngle"); break;
case ModelValue.XY: columns.Add("MoveAngle"); break;
default: throw new Exception("Unknown value for prediction : " + prediction);
}
}
return(columns.ToArray());
}
public void SetLevel1UnitValueTest()
{
//test Level 2
Model model = new Model();
ModelValue mv = new ModelValue();
SetModelValueSample(mv);
//calculate Level 3 values
mv.CalculateLevel4Values();
mv.CalculateLevel3Values();
//create model
model.CreateModel();
//GlobalData.GlobalData.globalData.if_assess_modularity = true;
//model.TrimModel();
model.CalculateModelWeight();
mv.SetLevel4UnitValue(model.level4_nodes);
mv.SetLevel3UnitValue(model.level3_nodes);
mv.SetLevel2UnitValue(model.level2_nodes);
mv.SetLevel1UnitValue(model.level1_nodes);
//verify the value
Assert.AreEqual(3.1944350372696791127913786374672, (double)model.level1_nodes[0].value, 0.0001);
////修剪模型的情况
//model.CreateModel();
//GlobalData.GlobalData.globalData.if_assess_modularity = false;
//model.TrimModel();
//model.CalculateModelWeight();
//mv.SetLevel4UnitValue(model.level4_nodes);
//mv.SetLevel3UnitValue(model.level3_nodes);
//mv.SetLevel2UnitValue(model.level2_nodes);
//mv.SetLevel1UnitValue(model.level1_nodes);
////验证数值
//Assert.AreEqual(2.8009873001760972650572709221841, (double)model.level1_nodes[0].value, 0.0001);
}
public void CalculateLevel3ValuesTest()
{
//build Model's Dictionary
ModelValue mv = new ModelValue();
SetModelValueSample(mv);
//calculate Level 3 values
mv.CalculateLevel4Values();
mv.CalculateLevel3Values();
//verify calculation results
Assert.AreEqual(mv.NDCS_TOTAL / mv.NS, mv.Coupling);
Assert.AreEqual(mv.NS / mv.NM, mv.Cohesion);
Assert.AreEqual(mv.NIS_TOTAL / (mv.NS * mv.NS - mv.NS), mv.Design_Complexity);
Assert.AreEqual((mv.NPI_TOTAL + mv.NII_TOTAL) / mv.NS, mv.System_Size);
Assert.AreEqual((mv.NPI_TOTAL * mv.NPI_TOTAL) / (mv.NM * mv.NM), mv.Service_Granularity);
Assert.AreEqual((mv.NMP_TOTAL * mv.NMP_TOTAL) / (mv.NPI_TOTAL * mv.NPI_TOTAL), mv.Parameter_Granularity);
Assert.AreEqual(mv.WISL_TOTAL / mv.NS, mv.Service_Loopback);
Assert.AreEqual(17.28380m / 12.23456m, mv.Coupling);
Assert.AreEqual(12.23456m / 23.45678m, mv.Cohesion);
Assert.AreEqual(42.71615m / (12.23456m * 12.23456m - 12.23456m), mv.Design_Complexity);
Assert.AreEqual((67.28380m + 92.71615m) / 12.23456m, mv.System_Size);
Assert.AreEqual((67.28380m * 67.28380m) / (23.45678m * 23.45678m), mv.Service_Granularity);
Assert.AreEqual((117.28380m * 117.28380m) / (67.28380m * 67.28380m), mv.Parameter_Granularity);
Assert.AreEqual(3 / 12.23456m, mv.Service_Loopback);
}
private static string WriteToDisk(IEnumerable <ModelDataSet> data, ModelValue prediction)
{
// geneate a random path
var path = Path.Combine(Path.GetTempPath(), Path.GetTempFileName());
using (var writer = File.CreateText(path))
{
foreach (var d in data)
{
for (int i = 0; i < d.Features(); i++)
{
writer.Write(d.Feature(i));
writer.Write(',');
}
switch (prediction)
{
case ModelValue.Action: writer.WriteLine(d.Action); break;
case ModelValue.Angle: writer.WriteLine(d.FaceAngle); break;
case ModelValue.XY: writer.WriteLine(d.MoveAngle); break;
default: throw new Exception("Unknown value for prediction : " + prediction);
}
}
}
return(path);
}
private static TextLoader GetTextLoader(MLContext context, ModelValue prediction)
{
var index = 0;
return(context.Data.CreateTextLoader(
columns: ColumnNames(true /*with label*/, prediction).Select(c => new TextLoader.Column(c, DataKind.Single, index++)).ToArray(),
separatorChar: ',',
hasHeader: false));
}
private void BindToModelValue(EAV.Model.IModelAttribute attribute, EAV.Model.IModelInstance instance, EAV.Model.IModelValue value, ViewModelAttributeValue viewValue)
{
if (value == null && viewValue == null)
{
return;
}
if (value != null && viewValue == null)
{
if (value.ObjectState == ObjectState.New)
{
instance.Values.Remove(value);
}
else
{
value.MarkDeleted();
}
return;
}
else if (value == null && viewValue != null)
{
value = new ModelValue()
{
Attribute = attribute, Instance = instance, Unit = attribute.VariableUnits.HasValue ? new ModelUnit()
{
UnitID = NextUnitID
} : null
};
}
if (value.RawValue != viewValue.Value)
{
value.RawValue = viewValue.Value;
}
if (attribute.VariableUnits.HasValue)
{
if (attribute.VariableUnits.Value)
{
if (value.Unit.ObjectState == ObjectState.New && !String.Equals(value.Unit.DisplayText, viewValue.UnitText, StringComparison.InvariantCulture))
{
value.Unit.DisplayText = viewValue.UnitText;
}
}
else
{
if (value.UnitID != viewValue.UnitID)
{
value.Unit = attribute.Units.SingleOrDefault(it => it.UnitID == viewValue.UnitID);
}
}
}
}
public void Execute(int s, ref int sP, ref int steps)
{
int lclsteps = 0;
sP = s;
double r_temp = 0;
//steps--;
int max = 5;
while (sP != SubGoalState)
{
int[] temp = maxA(sP);
int a = temp[1];
s = sP;
if (temp[0] == -1)
{
break;// theres no choice in initiation set from this state
}
//sP = temp[0];
bool contextChanged = m_agent.executeAction(sP, temp[1], ref sP, ref r_temp, ref steps);
if (sP == s)
{
max--;
if (max == 0)
{
break;
}
}
else
{
max = 5;
}
if (contextChanged || temp[0] != sP)
{
break;
}
if (m_agent.Model[sP, temp[1]] == null)
{
ModelValue m = new ModelValue(s, a, sP, r_temp, 1);
m_agent.Model[s, a] = m;
}
else
{
m_agent.Model[s, a].Update(s, a, sP, r_temp, 1);
}
//steps++;
lclsteps++;
if (lclsteps > 100)
{
break;
}
}
//steps++;
}
private static TextLoader GetTextLoader(MLContext context, ModelValue prediction)
{
var index = 0;
return(context.Data.TextReader(
new TextLoader.Arguments()
{
Separator = ",",
HasHeader = false,
Column = ColumnNames(true /*with label*/, prediction).Select(c => new TextLoader.Column(c, DataKind.R4, index++)).ToArray()
}));
}
// train
public ModelMLNet(IEnumerable <ModelDataSet> data, ModelValue prediction)
{
Context = new MLContext();
// add data
var textLoader = GetTextLoader(Context, prediction);
// spill to disk !?!?! since there is no way to load from a collection
var pathToData = "";
try
{
// write data to disk
pathToData = WriteToDisk(data, prediction);
// read in data
IDataView dataView = textLoader.Load(pathToData);
InputSchema = dataView.Schema;
// configurations
var label = "";
switch (prediction)
{
case ModelValue.Action: label = "Action"; break;
case ModelValue.Angle: label = "FaceAngle"; break;
case ModelValue.XY: label = "MoveAngle"; break;
default: throw new Exception("Unknown value for prediction : " + prediction);
}
var dataPipeline = Context.Transforms.CopyColumns(outputColumnName: "Label", inputColumnName: label)
.Append(Context.Transforms.Concatenate("Features", ColumnNames(false /* do not include the label */, prediction)));
// set the training algorithm
var trainer = Context.Regression.Trainers.Sdca(labelColumnName: "Label", featureColumnName: "Features");
var trainingPipeline = dataPipeline.Append(trainer);
TrainedModel = trainingPipeline.Fit(dataView);
}
finally
{
// cleanup
if (!string.IsNullOrWhiteSpace(pathToData) && File.Exists(pathToData))
{
File.Delete(pathToData);
}
}
}
public override ModelFitness Evaluate(List <ModelDataSet> data, ModelValue prediction)
{
if (TrainedModel == null)
{
throw new Exception("Must initialize the model before calling");
}
lock (TrainedModel)
{
#if ML_LEGACY
var testData = CollectionDataSource.Create(data);
var evaluator = new RegressionEvaluator();
var metrics = evaluator.Evaluate(TrainedModel, testData);
return(new ModelFitness()
{
RMS = metrics.Rms,
RSquared = metrics.RSquared
});
#else
var textLoader = GetTextLoader(Context, prediction);
var pathToData = "";
try
{
// ugh have to spill data to disk for it to work!
pathToData = WriteToDisk(data, prediction);
IDataView dataView = textLoader.Read(pathToData);
var predictions = TrainedModel.Transform(dataView);
var metrics = Context.Regression.Evaluate(predictions, label: "Label", score: "Score");
return(new ModelFitness()
{
RMS = metrics.Rms,
RSquared = metrics.RSquared
});
}
finally
{
// cleanup
if (!string.IsNullOrWhiteSpace(pathToData) && File.Exists(pathToData))
{
File.Delete(pathToData);
}
}
#endif
}
}
public void CalculateLevel4ValuesTest()
{
//build Model's Dictionary
ModelValue mv = new ModelValue();
SetModelValueSample(mv);
//calculate Level 4 values
mv.CalculateLevel4Values();
//verify calculation results
Assert.AreEqual(12.23456m, mv.NS);
Assert.AreEqual(17.28380m, mv.NDCS_TOTAL);
Assert.AreEqual(23.45678m, mv.NM);
Assert.AreEqual(42.71615m, mv.NIS_TOTAL);
Assert.AreEqual(67.28380m, mv.NPI_TOTAL);
Assert.AreEqual(92.71615m, mv.NII_TOTAL);
Assert.AreEqual(117.28380m, mv.NMP_TOTAL);
Assert.AreEqual(3, mv.WISL_TOTAL);
}
public void SetModelValueSample(ModelValue mv)
{
//add service name
mv.Dic_NO_ServiceName.Add(0, "Service 1");
mv.Dic_NO_ServiceName.Add(1, "Service 2");
mv.Dic_NO_ServiceName.Add(2, "Service 3");
mv.Dic_NO_ServiceName.Add(3, "Service 4");
mv.Dic_NO_ServiceName.Add(4, "Service 5");
//add number
mv.NS = 12.23456m;
mv.NDCS_NO_VALUE.Add(0, 1.23456m);
mv.NDCS_NO_VALUE.Add(1, 2.34567m);
mv.NDCS_NO_VALUE.Add(2, 3.45678m);
mv.NDCS_NO_VALUE.Add(3, 4.56789m);
mv.NDCS_NO_VALUE.Add(4, 5.67890m);
mv.NM = 23.45678m;
mv.NIS_NO_VALUE.Add(0, 6.78901m);
mv.NIS_NO_VALUE.Add(1, 7.89012m);
mv.NIS_NO_VALUE.Add(2, 8.90123m);
mv.NIS_NO_VALUE.Add(3, 9.01234m);
mv.NIS_NO_VALUE.Add(4, 10.12345m);
mv.NPI_NO_VALUE.Add(0, 11.23456m);
mv.NPI_NO_VALUE.Add(1, 12.34567m);
mv.NPI_NO_VALUE.Add(2, 13.45678m);
mv.NPI_NO_VALUE.Add(3, 14.56789m);
mv.NPI_NO_VALUE.Add(4, 15.67890m);
mv.NII_NO_VALUE.Add(0, 16.78901m);
mv.NII_NO_VALUE.Add(1, 17.89012m);
mv.NII_NO_VALUE.Add(2, 18.90123m);
mv.NII_NO_VALUE.Add(3, 19.01234m);
mv.NII_NO_VALUE.Add(4, 20.12345m);
mv.NMP_NO_VALUE.Add(0, 21.23456m);
mv.NMP_NO_VALUE.Add(1, 22.34567m);
mv.NMP_NO_VALUE.Add(2, 23.45678m);
mv.NMP_NO_VALUE.Add(3, 24.56789m);
mv.NMP_NO_VALUE.Add(4, 25.67890m);
mv.WISL_NO_VALUE.Add(0, 1);
mv.WISL_NO_VALUE.Add(1, 0);
mv.WISL_NO_VALUE.Add(2, 1);
mv.WISL_NO_VALUE.Add(3, 1);
mv.WISL_NO_VALUE.Add(4, 0);
}
public void SetLevel2UnitValueTest()
{
//test Level 2
Model model = new Model();
ModelValue mv = new ModelValue();
SetModelValueSample(mv);
//calculate Level 3 values
mv.CalculateLevel4Values();
mv.CalculateLevel3Values();
//create model
model.CreateModel();
//GlobalData.GlobalData.globalData.if_assess_modularity = true;
//model.TrimModel();
model.CalculateModelWeight();
mv.SetLevel4UnitValue(model.level4_nodes);
mv.SetLevel3UnitValue(model.level3_nodes);
mv.SetLevel2UnitValue(model.level2_nodes);
//verify the value
Assert.AreEqual(-0.5814090208767396425967497017122, (double)model.level2_nodes[0].value, 0.0001);
Assert.AreEqual(4.5493525993415009929712588253279, (double)model.level2_nodes[1].value, 0.0001);
Assert.AreEqual(4.435102352522840320814015528183, (double)model.level2_nodes[2].value, 0.0001);
Assert.AreEqual(4.37469421809111477997698989807, (double)model.level2_nodes[3].value, 0.0001);
////修剪模型的情况
//model.CreateModel();
//GlobalData.GlobalData.globalData.if_assess_modularity = false;
//model.TrimModel();
//model.CalculateModelWeight();
//mv.SetLevel4UnitValue(model.level4_nodes);
//mv.SetLevel3UnitValue(model.level3_nodes);
//mv.SetLevel2UnitValue(model.level2_nodes);
////验证数值
//Assert.AreEqual(-0.5814090208767396425967497017122, (double)model.level2_nodes[0].value, 0.0001);
//Assert.AreEqual(4.5493525993415009929712588253279, (double)model.level2_nodes[1].value, 0.0001);
//Assert.AreEqual(4.435102352522840320814015528183, (double)model.level2_nodes[2].value, 0.0001);
}
public override ModelFitness Evaluate(List <ModelDataSet> data, ModelValue prediction)
{
if (TrainedModel == null)
{
throw new Exception("Must initialize the model before calling");
}
lock (TrainedModel)
{
var textLoader = GetTextLoader(Context, prediction);
var pathToData = "";
try
{
// ugh have to spill data to disk for it to work!
pathToData = WriteToDisk(data, prediction);
IDataView dataView = textLoader.Load(pathToData);
var predictions = TrainedModel.Transform(dataView);
var metrics = Context.Regression.Evaluate(predictions, labelColumnName: "Label", scoreColumnName: "Score");
return(new ModelFitness()
{
RMS = metrics.RootMeanSquaredError,
RSquared = metrics.RSquared
});
}
finally
{
// cleanup
if (!string.IsNullOrWhiteSpace(pathToData) && File.Exists(pathToData))
{
File.Delete(pathToData);
}
}
}
}
private void SetDpItemValue(IWriteSession session, IDpItem dpItem, ModelValue value)
{
dpItem.Value = value;
session.SaveHierarchy(dpItem);
}
public override ModelFitness Evaluate(List <ModelDataSet> data, ModelValue prediction)
{
if (TrainedModel == null)
{
throw new Exception("Must initialize the model before calling");
}
if (data.Count == 0)
{
return(default(ModelFitness));
}
// values
var predictions = new float[data.Count];
var labels = new float[data.Count];
for (int i = 0; i < data.Count; i++)
{
predictions[i] = Predict(data[i]);
switch (prediction)
{
case ModelValue.Action: labels[i] = (float)data[i].Action; break;
case ModelValue.Angle: labels[i] = data[i].FaceAngle; break;
case ModelValue.XY: labels[i] = data[i].MoveAngle; break;
default: throw new Exception("Unknown prediction type : " + prediction);
}
}
// calculate rms
// https://sciencing.com/calculate-rms-5104500.html
// rms(a,b,c) = sqrt((a^2 + b^2 + c^2)/3);
var rms = 0d;
for (int i = 0; i < predictions.Length; i++)
{
rms = Math.Pow(predictions[i] - labels[i], 2);
}
rms /= predictions.Length;
rms = Math.Sqrt(rms);
// calculate r^2
// https://www.sapling.com/5117357/calculate-rsquared
// R = (Count (sum of ab) - (sum of a)(sum of b)) / [sqrt((Count(sum a^2) - (sum of a)^2)(Count *(sum of b^2) - (sum of b)^2)]
// a == labels
// b == prediction
var ab = 0d;
var a = 0d;
var b = 0d;
var a2 = 0d;
var b2 = 0d;
for (int i = 0; i < predictions.Length; i++)
{
a += labels[i];
b += predictions[i];
ab += (labels[i] * predictions[i]);
a2 += Math.Pow(labels[i], 2);
b2 += Math.Pow(predictions[i], 2);
}
var r2 = ((predictions.Length * ab) - (a * b)) / Math.Sqrt(((predictions.Length * a2) - Math.Pow(a, 2)) * (predictions.Length * b2 - Math.Pow(b, 2)));
r2 = Math.Pow(r2, 2);
return(new ModelFitness()
{
RMS = rms,
RSquared = r2
});
}
public void SetLevel3UnitValueTest()
{
Model model = new Model();
ModelValue mv = new ModelValue();
SetModelValueSample(mv);
//calculate Level 3 values
mv.CalculateLevel4Values();
mv.CalculateLevel3Values();
model.CreateModel();
//GlobalData.GlobalData.globalData.if_assess_modularity = true;
//model.TrimModel();
model.CalculateModelWeight();
mv.SetLevel4UnitValue(model.level4_nodes);
mv.SetLevel3UnitValue(model.level3_nodes);
mv.SetLevel3UnitValue(model.level2_nodes);
mv.SetLevel3UnitValue(model.level1_nodes);
//verify the value of Level 3
Assert.AreEqual(17.28380m / 12.23456m, model.level3_nodes[0].value);
Assert.AreEqual(12.23456m / 23.45678m, model.level3_nodes[1].value);
Assert.AreEqual(42.71615m / (12.23456m * 12.23456m - 12.23456m), model.level3_nodes[2].value);
Assert.AreEqual((67.28380m + 92.71615m) / 12.23456m, model.level3_nodes[3].value);
Assert.AreEqual(17.28380m / 12.23456m, model.level3_nodes[4].value);
Assert.AreEqual(12.23456m / 23.45678m, model.level3_nodes[5].value);
Assert.AreEqual((67.28380m * 67.28380m) / (23.45678m * 23.45678m), model.level3_nodes[6].value);
Assert.AreEqual((117.28380m * 117.28380m) / (67.28380m * 67.28380m), model.level3_nodes[7].value);
Assert.AreEqual(3 / 12.23456m, model.level3_nodes[8].value);
Assert.AreEqual((117.28380m * 117.28380m) / (67.28380m * 67.28380m), model.level3_nodes[9].value);
Assert.AreEqual(3 / 12.23456m, model.level3_nodes[10].value);
Assert.AreEqual(12.23456m / 23.45678m, model.level3_nodes[11].value);
Assert.AreEqual((67.28380m * 67.28380m) / (23.45678m * 23.45678m), model.level3_nodes[12].value);
//verify the value of Level 2
foreach (var node in model.level2_nodes)
{
Assert.AreEqual(0, node.value);
}
//verify the value of Level 1
foreach (var node in model.level1_nodes)
{
Assert.AreEqual(0, node.value);
}
////裁剪模型的情况
//model.CreateModel();
//GlobalData.GlobalData.globalData.if_assess_modularity = false;
//model.TrimModel();
//model.CalculateModelWeight();
//mv.SetLevel4UnitValue(model.level4_nodes);
//mv.SetLevel3UnitValue(model.level3_nodes);
//mv.SetLevel3UnitValue(model.level2_nodes);
//mv.SetLevel3UnitValue(model.level1_nodes);
////验证第3层的值
//Assert.AreEqual(17.28380m / 12.23456m, model.level3_nodes[0].value);
//Assert.AreEqual(12.23456m / 23.45678m, model.level3_nodes[1].value);
//Assert.AreEqual(42.71615m / (12.23456m * 12.23456m - 12.23456m), model.level3_nodes[2].value);
//Assert.AreEqual((67.28380m + 92.71615m) / 12.23456m, model.level3_nodes[3].value);
//Assert.AreEqual(17.28380m / 12.23456m, model.level3_nodes[4].value);
//Assert.AreEqual(12.23456m / 23.45678m, model.level3_nodes[5].value);
//Assert.AreEqual((67.28380m * 67.28380m) / (23.45678m * 23.45678m), model.level3_nodes[6].value);
//Assert.AreEqual((117.28380m * 117.28380m) / (67.28380m * 67.28380m), model.level3_nodes[7].value);
//Assert.AreEqual(3 / 12.23456m, model.level3_nodes[8].value);
//Assert.AreEqual((117.28380m * 117.28380m) / (67.28380m * 67.28380m), model.level3_nodes[9].value);
//Assert.AreEqual(3 / 12.23456m, model.level3_nodes[10].value);
////验证第2层的值
//foreach (var node in model.level2_nodes)
// Assert.AreEqual(0, node.value);
////验证第1层的值
//foreach (var node in model.level1_nodes)
// Assert.AreEqual(0, node.value);
}
public virtual ModelFitness Evaluate(List <ModelDataSet> data, ModelValue prediction)
{
return(default(ModelFitness));
}
// train
public Model(IEnumerable <ModelDataSet> data, ModelValue prediction)
{
}
// train
public ModelMLNet(IEnumerable <ModelDataSet> data, ModelValue prediction)
{
#if ML_LEGACY
var pipeline = new LearningPipeline();
// add data
pipeline.Add(CollectionDataSource.Create(data));
// choose what to predict
switch (prediction)
{
case ModelValue.Action:
pipeline.Add(new ColumnCopier(("Action", "Label")));
break;
case ModelValue.Angle:
pipeline.Add(new ColumnCopier(("FaceAngle", "Label")));
break;
case ModelValue.XY:
pipeline.Add(new ColumnCopier(("MoveAngle", "Label")));
break;
default: throw new Exception("Unknown prediction : " + prediction);
}
// add columns as features
// do not include the features which should be predicted
pipeline.Add(new ColumnConcatenator("Features",
"CenterAngle",
"InZone",
"Health",
"Shield",
"Primary",
"PrimaryAmmo",
"PrimaryClip",
"Secondary",
"SecondaryAmmo",
"SecondaryClip",
// "Ammo",
"Angle_1",
"Distance_1",
// "Bandage",
"Angle_2",
"Distance_2",
// "Helmet",
"Angle_3",
"Distance_3",
// "Ak47",
"Angle_4",
"Distance_4",
// "Shotgun",
"Angle_5",
"Distance_5",
// "Pistol",
"Angle_6",
"Distance_6",
// "Obstacle",
"Angle_7",
"Distance_7",
// "Player",
"Angle_8",
"Distance_8"
));
// add a classifier
// action 0.25, xy 84.9777, angle 71.591
//pipeline.Add(new FastTreeRegressor());
pipeline.Add(new FastTreeRegressor()
{
// NumTrees
// 100 - default
// 100 - xy 79
// 1000 - xy 63, angle 52 (slow)
NumTrees = 100,
// NumLeaves
// 50 - default
// 100 - xy 63, angle 71
// 1000 - xy 59, angle 49 (slow)
NumLeaves = 50,
// NumThreads
// 5 - default
NumThreads = 50,
// EntropyCoe
// 0.30 - default
// 0.70 - xy 79, angle 66
// 0.99 - xy 79, angle 66
// 0.05 - xy 79, angle 66
EntropyCoefficient = 0.3
});
// action 0.26 xy 85.1606, angle 72.5194
//pipeline.Add(new FastTreeTweedieRegressor());
// took too long
//pipeline.Add(new GeneralizedAdditiveModelRegressor());
// runtime exception
//pipeline.Add(new LightGbmRegressor());
// action 0.4736, xy 105.8815, angle 91.2677
//pipeline.Add(new OnlineGradientDescentRegressor());
// runtime exception
//pipeline.Add(new OrdinaryLeastSquaresRegressor());
// action 0.4628, xy 106.0547, angle 91.0179
//pipeline.Add(new PoissonRegressor());
// action 0.4599, xy 105.8474, angle 90.4134
//pipeline.Add(new StochasticDualCoordinateAscentRegressor());
// train the model
TrainedModel = pipeline.Train <ModelDataSet, ModelDataSetPrediction>();
#else
Context = new MLContext();
// add data
var textLoader = GetTextLoader(Context, prediction);
// spill to disk !?!?! since there is no way to load from a collection
var pathToData = "";
try
{
// write data to disk
pathToData = WriteToDisk(data, prediction);
// read in data
IDataView dataView = textLoader.Read(pathToData);
// configurations
var label = "";
switch (prediction)
{
case ModelValue.Action: label = "Action"; break;
case ModelValue.Angle: label = "FaceAngle"; break;
case ModelValue.XY: label = "MoveAngle"; break;
default: throw new Exception("Unknown value for prediction : " + prediction);
}
var dataPipeline = Context.Transforms.CopyColumns(label, "Label")
.Append(Context.Transforms.Concatenate("Features", ColumnNames(false /* do not include the label */, prediction)));
// set the training algorithm
var trainer = Context.Regression.Trainers.FastTree(label: "Label", features: "Features");
var trainingPipeline = dataPipeline.Append(trainer);
TrainedModel = trainingPipeline.Fit(dataView);
}
finally
{
// cleanup
if (!string.IsNullOrWhiteSpace(pathToData) && File.Exists(pathToData))
{
File.Delete(pathToData);
}
}
#endif
}
public void SetLevel4UnitValueTest()
{
Model model = new Model();
ModelValue mv = new ModelValue();
SetModelValueSample(mv);
//calculate Level 4 values
mv.CalculateLevel4Values();
model.CreateModel();
//GlobalData.GlobalData.globalData.if_assess_modularity = true;
//model.TrimModel();
model.CalculateModelWeight();
mv.SetLevel4UnitValue(model.level4_nodes);
mv.SetLevel4UnitValue(model.level3_nodes);
mv.SetLevel4UnitValue(model.level2_nodes);
mv.SetLevel4UnitValue(model.level1_nodes);
//verify the value of Level 4
Assert.AreEqual(12.23456m, model.level4_nodes[0].value);
Assert.AreEqual(17.28380m, model.level4_nodes[1].value);
Assert.AreEqual(12.23456m, model.level4_nodes[2].value);
Assert.AreEqual(23.45678m, model.level4_nodes[3].value);
Assert.AreEqual(12.23456m, model.level4_nodes[4].value);
Assert.AreEqual(42.71615m, model.level4_nodes[5].value);
Assert.AreEqual(12.23456m, model.level4_nodes[6].value);
Assert.AreEqual(67.28380m, model.level4_nodes[7].value);
Assert.AreEqual(92.71615m, model.level4_nodes[8].value);
Assert.AreEqual(12.23456m, model.level4_nodes[9].value);
Assert.AreEqual(17.28380m, model.level4_nodes[10].value);
Assert.AreEqual(12.23456m, model.level4_nodes[11].value);
Assert.AreEqual(23.45678m, model.level4_nodes[12].value);
Assert.AreEqual(12.23456m, model.level4_nodes[13].value);
Assert.AreEqual(23.45678m, model.level4_nodes[14].value);
Assert.AreEqual(67.28380m, model.level4_nodes[15].value);
Assert.AreEqual(23.45678m, model.level4_nodes[16].value);
Assert.AreEqual(117.28380m, model.level4_nodes[17].value);
Assert.AreEqual(67.28380m, model.level4_nodes[18].value);
Assert.AreEqual(12.23456m, model.level4_nodes[19].value);
Assert.AreEqual(3, model.level4_nodes[20].value);
Assert.AreEqual(23.45678m, model.level4_nodes[21].value);
Assert.AreEqual(117.28380m, model.level4_nodes[22].value);
Assert.AreEqual(67.28380m, model.level4_nodes[23].value);
Assert.AreEqual(12.23456m, model.level4_nodes[24].value);
Assert.AreEqual(3, model.level4_nodes[25].value);
Assert.AreEqual(12.23456m, model.level4_nodes[26].value);
Assert.AreEqual(23.45678m, model.level4_nodes[27].value);
Assert.AreEqual(12.23456m, model.level4_nodes[28].value);
Assert.AreEqual(23.45678m, model.level4_nodes[29].value);
Assert.AreEqual(67.28380m, model.level4_nodes[30].value);
//verify the value of Level 3
foreach (var node in model.level3_nodes)
{
Assert.AreEqual(0, node.value);
}
//verify the value of Level 2
foreach (var node in model.level2_nodes)
{
Assert.AreEqual(0, node.value);
}
//verify the value of Level 1
foreach (var node in model.level1_nodes)
{
Assert.AreEqual(0, node.value);
}
////裁剪模型的情况
//model.CreateModel();
//GlobalData.GlobalData.globalData.if_assess_modularity = false;
//model.TrimModel();
//model.CalculateModelWeight();
//mv.SetLevel4UnitValue(model.level4_nodes);
//mv.SetLevel4UnitValue(model.level3_nodes);
//mv.SetLevel4UnitValue(model.level2_nodes);
//mv.SetLevel4UnitValue(model.level1_nodes);
////验证第4层的值
//Assert.AreEqual(12.23456m, model.level4_nodes[0].value);
//Assert.AreEqual(17.28380m, model.level4_nodes[1].value);
//Assert.AreEqual(12.23456m, model.level4_nodes[2].value);
//Assert.AreEqual(23.45678m, model.level4_nodes[3].value);
//Assert.AreEqual(12.23456m, model.level4_nodes[4].value);
//Assert.AreEqual(42.71615m, model.level4_nodes[5].value);
//Assert.AreEqual(12.23456m, model.level4_nodes[6].value);
//Assert.AreEqual(67.28380m, model.level4_nodes[7].value);
//Assert.AreEqual(92.71615m, model.level4_nodes[8].value);
//Assert.AreEqual(12.23456m, model.level4_nodes[9].value);
//Assert.AreEqual(17.28380m, model.level4_nodes[10].value);
//Assert.AreEqual(12.23456m, model.level4_nodes[11].value);
//Assert.AreEqual(23.45678m, model.level4_nodes[12].value);
//Assert.AreEqual(12.23456m, model.level4_nodes[13].value);
//Assert.AreEqual(23.45678m, model.level4_nodes[14].value);
//Assert.AreEqual(67.28380m, model.level4_nodes[15].value);
//Assert.AreEqual(23.45678m, model.level4_nodes[16].value);
//Assert.AreEqual(117.28380m, model.level4_nodes[17].value);
//Assert.AreEqual(67.28380m, model.level4_nodes[18].value);
//Assert.AreEqual(12.23456m, model.level4_nodes[19].value);
//Assert.AreEqual(3, model.level4_nodes[20].value);
//Assert.AreEqual(23.45678m, model.level4_nodes[21].value);
//Assert.AreEqual(117.28380m, model.level4_nodes[22].value);
//Assert.AreEqual(67.28380m, model.level4_nodes[23].value);
//Assert.AreEqual(12.23456m, model.level4_nodes[24].value);
//Assert.AreEqual(3, model.level4_nodes[25].value);
////验证第3层的值
//foreach (var node in model.level3_nodes)
// Assert.AreEqual(0, node.value);
////验证第2层的值
//foreach (var node in model.level2_nodes)
// Assert.AreEqual(0, node.value);
////验证第1层的值
//foreach (var node in model.level1_nodes)
// Assert.AreEqual(0, node.value);
}
public Value(ModelValue model, double value = 0)
{
this.model = model;
this.value = value;
}
public string GetRawModelXML(ModelValue model)
{
return((model?.NameVersion?.CCacheId as NodeSetFileModel)?.FileCache);
}
protected ModelValue DetermineParameterValue(ModelParameter parameter)
{
ModelValues values = new ModelValues();
values.Add(parameter.Values);
//Parameter values, can be specified in numerous ways:
// 1. Value list - simply choose one of them
// 2. Expression - generate value that meets the criteria (ie: <, >, !=, etc)
// 3. Variable - simply obtain the value by calling a method/field
//#3. Variable - simply obtain the value by calling a method/field
if(parameter.Variable != null)
{
object current = parameter.Variable.CachedValue;
if(current is IEnumerable && !typeof(IEnumerable).IsAssignableFrom(parameter.Type))
{
foreach(object v in (IEnumerable)current)
values.Add(new ModelValue(v));
}
else
{
values.Add(new ModelValue(current));
}
}
//First ensure we have a set of values/requirements to choose from
if(values.Count <= 0)
throw new ModelException(parameter, "No values specified to choose from");
//Note: Since we allow the operator on the individual values, this is a little more complex.
//Note: We allow multiple operators, not just one, (ie: x > 5, < 10, and != 7).
//This gives you great power and flexibility in expressions, but means we have to work a
//little hard in determing a value that meets the requirements
//#1. Value list - simply choose one of them
//Note: Bitmask is already exploded into combinations
ModelValues equalvalues = values.FindOperator(ModelValueOperator.Equal);
//#2. Expression - generate value that meets the criteria (ie: <, >, !=, etc)
//Note: Since we allow the operator on the individual values, this is a little more complex.
//Note: We allow multiple operators, not just one, (ie: x > 5, < 10, and != 7).
//This gives you great power and flexibility in expressions, but means we have to work a
//little hard in determing a value that meets the requirements.
int min = Int32.MinValue;
int max = Int32.MaxValue;
//Adjust our parameter, simplier to loop over all of them
foreach(ModelValue value in values.FindOperator(ModelValueOperator.Equal, false).FindOperator(ModelValueOperator.NotEqual, false))
{
//To keep this simple, we just support integers (for now).
if(!(value.Value is int || value.Value is Nullable<int>))
throw new ModelException(parameter, "Generated value range must be specified in terms of integers, not '" + value.Type + "'");
//Simplify our life
int v = (int)value.Value;
//Adjust Max (if there is one)
switch(value.Operator)
{
case ModelValueOperator.LessThanOrEqual:
if(v < max)
max = v;
break;
case ModelValueOperator.GreaterThanOrEqual:
if(v > min)
min = v;
break;
case ModelValueOperator.LessThan:
if(v-1 < max && v > Int32.MinValue/*prevent underflow*/)
max = v-1;
break;
case ModelValueOperator.GreaterThan:
if(v+1 > min && v < Int32.MaxValue/*prevent overflow*/)
min = v+1;
break;
};
}
//Choose a new value, within the specified range
//Note: We retry in case it equals one of the existing invalid values (ie: !=)
while(true)
{
ModelValue choice = null;
if(equalvalues.Count > 0)
{
//Simple: Choose one of the specified values
choice = new ModelValue(equalvalues.Choose(this).Value);
}
else
{
//Otherwise: Choose a value within in the range
//Note: Random.Next = min <= x < max (so we have to add one)
int index = _options.Random.Next(min, max < Int32.MaxValue ? max + 1 : max); //Prevent overflow
choice = new ModelValue(index);
}
//As soon as we find a value, within the range, (and not in the invalid list), were done
bool valid = true;
foreach(ModelValue invalid in values.FindOperator(ModelValueOperator.NotEqual))
{
if(invalid.Evaluate(choice.Value, ModelValueOperator.Equal))
valid = false;
}
if(valid)
return choice;
}
}
// train
public ModelOpenCV(List <ModelDataSet> input, ModelValue prediction)
{
if (input == null || input.Count == 0)
{
throw new Exception("Must have valid input");
}
// convert features into proper form
var features = new float[input.Count, input[0].Features()];
var labels = new float[input.Count];
for (int i = 0; i < input.Count; i++)
{
for (int j = 0; j < input[i].Features(); j++)
{
features[i, j] = input[i].Feature(j);
}
switch (prediction)
{
case ModelValue.Action: labels[i] = input[i].Action; break;
case ModelValue.Angle: labels[i] = input[i].FaceAngle; break;
case ModelValue.XY: labels[i] = input[i].MoveAngle; break;
default: throw new Exception("Unknown prediction type : " + prediction);
}
}
// train
var labelInput = InputArray.Create <float>(labels);
var dataInput = InputArray.Create <float>(features);
TrainedModel = RTrees.Create();
TrainedModel.RegressionAccuracy = 0.00001f;
// RTrees.MaxDepath (r^2)
// default - action 0.3424, xy 0.1735, angle 0.2208
// 5 -
// 20 - action 0.6482, xy 0.5912, angle 0.6414 (new default)
// 100 - action 0.6408, xy 0.5914, angle 0.6419
TrainedModel.MaxDepth = 20;
// RTress.MinSampleCount
// default(10) - see 20 above
// 1 - actions 0.6625, xy 0.5077, angle 0.6376
// 50 - actions 0.6464, xy 0.5627, angle 0.6217
//TrainedModel.MinSampleCount = 1;
// fails
//TrainedModel = LogisticRegression.Create();
// fails
// TrainedModel = DTrees.Create();
// failed
//TrainedModel = SVM.Create();
//TrainedModel.KernelType = SVM.KernelTypes.Linear;
//TrainedModel.Type = SVM.Types.NuSvr;
//TrainedModel.C = 1;
//TrainedModel.P = 0.01;
//TrainedModel.Gamma = 10f;
//TrainedModel.Degree = 0.1;
//TrainedModel.Coef0 = 0;
//TrainedModel.Nu = 0.1;
TrainedModel.Train(dataInput, SampleTypes.RowSample, labelInput);
}
//Constructor
public ModelExpression(ModelFunction func, ModelValue value)
: base(null, value)
{
_func = func;
}
public void LearnPS(int N, int episodes)
{
subgoalVar = new List <double>();
forceStop = false;
//int stateHold = currentState;
float alpha = 0.2f, gamma = 0.95f, theta = 0.001f;
int s = currentState; //Initialize
List <PriorityObj> PQueue = new List <PriorityObj>();
visit = new Dictionary <int, int>();
int epsilon_counter = 0;
int steps = 0;
while (true)
{
if (forceStop)
{
break;
}
if (epsilon_counter++ % 100 == 0)
{
epsilon_counter = 1;
e = 0.01 + 0.7 * Math.Pow(Math.E, -(discount++) / 20.0);
}
int tempSteps = steps; //Number of steps per episode
if (EpisodeSteps.Count > 1500)
{
forceStop = true;
}
//(b) a <- policy(s,Q)
//int a = e_greedy(Q[s]);
int a = softmaxWVisits(Q[s]);
// Q[s][a] -= -0.001;
double tempVal = Q[s][a];
//tempFirstVisitHist[s[0], s[1]] += 1;
//(c) Execute action a
int sP = 0;
double r = 0;
Context curCtxt = currentContext;
lock (threadLock)
{
executeAction(s, a, ref sP, ref r, ref steps);
}
if (!visit.ContainsKey(sP))
{
visit.Add(sP, 0);
}
Context nexCtxt = currentContext;
if (curCtxt != nexCtxt)
{
//double tmp = curCtxt.Q[s][Q[s].GetMaxAction()];
//for (int i = 0; i < curCtxt.Q[s].Actions.Count; i++)
//{
// Q[s][i] += curCtxt.Q[s][i];//tmp / 8;
//}
continue;
}
//SwitchContext(curCtxt);
int result = sP;
//(d) Model(s,a) <- s', r
//Model[s[0], s[1], a] = new int[] { sP[0], sP[1], r, steps - tempSteps };
if (Model[s, a] == null)
{
ModelValue m = new ModelValue(s, a, sP, r, steps - tempSteps);
Model[s, a] = m;
}
else
if (Model[s, a].calculateEm(sP, r) > 0)
{
Model[s, a].Update(s, a, sP, r, steps - tempSteps);
}
//System.Diagnostics.Debug.WriteLine("\tContext: " + currentContext.cID + "\t(s,a,sp): " + s + ", " + a + ", " + sP + "\tTm0: " /*+ (Model[s, a].Tm.Count > 0 ? Model[s,a].Tm[0].ToString() : " ") */ + "\tem0: " + em + "\tEm0: " + contexts[0].Em + "\tEm1: " + (contexts.Count > 1 ? contexts[1].Em.ToString() : "-"));
//(e) p <- |r + gama*maxA'Q(s',a') - Q(s,a)|
//float p = Math.Abs(r + gamma * getQ(sP[0], sP[1], maxA(sP)) - getQ(s[0], s[1], a));
double p = Math.Abs(r + gamma * Q[sP][Q[sP].GetMaxAction()] - Q[s][a]);
//(f) if p > tetha, then insert s,a into PQueue with priority p
PQueue.Clear();
if (p > theta)
{
InsertQueue(PQueue, s, a, p);
}
//(g) Repeat N times while PQueue is not empty
for (int i = 0; i < N && 0 < PQueue.Count; i++)
{
//(-)s, a <- first(PQueue)
PriorityObj obj = PQueue[0];// dene!!!!!!!!!!!!11
PQueue.Remove(obj);
s = obj.State;
a = obj.Action;
//(-)s', r <- Modell(s,a)
sP = Model[s, a].sP;
r = Model[s, a].reward;
int tsteps = Model[s, a].steps;
//(-)Q(s, a) <- Q(s,a) + alpha[r + gama*maxA'Q(s',a') - Q(s,a)]
//e_updates.Add(new double[] { s, a, Q[s][a] });
Q[s][a] = Q[s][a] + alpha * (r + Math.Pow(gamma, tsteps) * Q[sP][Q[sP].GetMaxAction()] - Q[s][a]);
//float t = (float)(getQ(s[0], s[1], a) + alpha * (r + Math.Pow(gamma, tsteps) * getQ(sP[0], sP[1], maxA(sP)) - getQ(s[0], s[1], a)));
//setQ(s[0], s[1], a, t);
//(-)Repeat, for all s",a" predicted to lead to s
List <ModelValue> list = Model.All_SA_PredictedToLeadTo(s);
for (int j = 0; j < list.Count; j++)
{
int sDP = list[j].s;
int aDP = list[j].a;
// r" <- predicted reward
double rDP = list[j].reward;
// p <- |r" + gamma*maxaQ(s,a) - Q(s",a")|
p = Math.Abs(rDP + gamma * Q[s][Q[s].GetMaxAction()] - Q[sDP][aDP]);
// if p > tetha, then insert s",a" into PQueue with priority p
if (p > theta)
{
InsertQueue(PQueue, sDP, aDP, p);
}
}
}
s = result;
environment.State = s;
if (environment.isGoal(s))
{
//visit = new Dictionary<int, int>();///////////
//if (subgoalValues != null)
//{
// subgoalValues[s] = 0.1;
// for (int f = 0; f < 4; f++)
// {
// if (Model.StateCount < s)
// Model.States.Add(new ModelState(s));
// if (Model[s] == null)
// Model[s] = new ModelState(s);
// if(Model[s].Count <= 4)
// Model[s].Add(new ModelValue(s, f, s, 0, 0));
// }
//}
if (Model.States.Count <= s)
{
Model[s, a] = null;
Model.States[s] = new ModelState(s);
}
//e = 0.9 * Math.Pow(Math.E, -(discount++) / 50.0); //Update epsilon
if (!DisableOptions && !m_optsLearned)
{
//initiationSetMembers = new int[Model.StateCount];
//initiationSets = new List<InitiationSet>();
Subgoals = SearchSubgoals();
//Subgoals.Add(new SubGoal(s));
SubgoalCounts.Add(Subgoals.Count);
double mean = 0;
double std = 100;
int sc = 10;
if (SubgoalCounts.Count > sc)
{
std = 0;
for (int c = SubgoalCounts.Count - sc; c < SubgoalCounts.Count; c++)
{
mean += SubgoalCounts[c] / (double)sc;
}
for (int c = SubgoalCounts.Count - sc; c < SubgoalCounts.Count; c++)
{
std += Math.Pow(SubgoalCounts[c] - mean, 2) / sc;
}
subgoalVar.Add(std);
std = Math.Sqrt(std);
}
else
{
subgoalVar.Add(-1);
}
if (std < 0.01)
{
List <OptionN> options = createOptFromSubGoals(Subgoals);
currentContext.options = options;
optNonLearnedVals = 0;
optLearnedVals = 0;
for (int i = 0; i < Model.StateCount; i++)
{
if (Model[i] == null)
{
optNonLearnedVals += 4;
}
else
{
for (int j = 0; j < Model[i].Count; j++)
{
if (Model[i][j] == null || Q[i][j] == 0)
{
optNonLearnedVals += 1;
}
else
{
optLearnedVals += 1;
}
}
}
}
//SearchInitiationSets(Subgoals); //Bu satır if in içinde olmalı (ki öyle(değil mi?))
//List<Option> options = CreateOptions();
OptionLearn(options);
//currentContext.options = options;
}
}
s = currentState;
EpisodeSteps.Add(steps);
ContextTrack.Add(currentContext.cID);
currentContext.EpisodeSteps.Add(steps);
steps = 0;
int calcLength = 9;
if (currentContext.EpisodeSteps.Count > calcLength)
{
double mean = 0;
double std = 0;
for (int c = currentContext.EpisodeSteps.Count - calcLength; c < currentContext.EpisodeSteps.Count; c++)
{
mean += EpisodeSteps[c] / (double)calcLength;
}
for (int c = currentContext.EpisodeSteps.Count - calcLength; c < currentContext.EpisodeSteps.Count; c++)
{
std += Math.Pow(currentContext.EpisodeSteps[c] - mean, 2) / calcLength;
}
std = Math.Sqrt(std);
currentContext.std = std;
bool stop = true;
for (int cx = 0; cx < contexts.Count; cx++)
{
if (contexts[cx].std > 10)
{
stop = false;
}
//System.Diagnostics.Debug.Write(cx + ":" + contexts[cx].std + " ");
}
//System.Diagnostics.Debug.Write("\n");
if (stopByEpisode)
{
if (stopEpisode <= EpisodeSteps.Count)
{
break;
}
}
else
{
if (stop || forceStop)
{
break;
}
}
}//Varyans ile dene
//if (!m_optsLearned)
//{
// OptionSearch();
// OptionLearn();
//}
PQueue.Clear();
if (curCtxt != nexCtxt)
{
SwitchContext(nexCtxt);
}
if (oneStepLearn)
{
break;
}
e_updates = new List <double[]>();
}
}
environment.State = currentState;
}
public DataAnalyze(JSONDataStruct.Rootobject datasource, ModelValue mv)
{
data = datasource;
modelValue = mv;
}
public static Value Max(ModelValue model)
{
return(new Value(model, model.Max));
}
public void Add(ModelValue value)
{
modelValues.Add(value);
}
