private static StringBuilder RecursiveTreePrintBuilder(DecisionTree.DecisionNode root, StringBuilder output, string indentation = "", int level = 0)
{
if (output == null)
{
output = new StringBuilder();
}
else
{
indentation += "\t";
}
if (root.Actor == null)
{
output.AppendLine("ROOT");
indentation += "\t";
}
if (root.DaughterNodes == null)
{
indentation += "\t";
output.AppendLine(indentation + "Leaf " + root.FeatureType + " " + root.Operator + " " + root.incomingEgde + " " + root.Action.StateName + " " + root.Actor);
}
else
{
foreach (DecisionTree.DecisionNode d in root.DaughterNodes)
{
output.AppendLine(indentation + Array.IndexOf(root.DaughterNodes, d).ToString() + " " + d.FeatureType + " " + d.Operator + " " + d.incomingEgde + " " + d.Action.StateName + " " + d.Actor);
RecursiveTreePrintBuilder(d, output, indentation, ++level);
}
}
return(output);
}
public DecisionTree(TrainingExample[] examples, Feature[] features, DecisionNode decisionNode)
{
rootNode = decisionNode;
Features = features;
TreeLearner(examples, features, decisionNode);
}
public void DecisionTreePrinter(DecisionNode startNode)
{
// Console.WriteLine("Feature for Root is " + startNode.TestValue.ID.ToString());
if (startNode.DaughterNodes != null)
{
for (int i = 0; i < startNode.DaughterNodes.Length; i++)
{
if (startNode.DaughterNodes[i] != null)
{
// UnityEngine.Debug.Log("Decision Node " + startNode.TestValue.ID + " Incoming Egde " + startNode.DaughterNodes[i].incomingEgde);
DecisionTreePrinter(startNode.DaughterNodes[i]);
}
}
}
else
{
if (startNode.Probability != null)
{
Debug.Log("Node: " + startNode.ToString());
foreach (var pair in startNode.Probability)
{
Debug.Log("Probability of State: " + pair.Key + " : " + pair.Value);
}
}
}
}
public DecisionNode(
Question question = null,
DecisionNode trueBranch = null,
DecisionNode falseBranch = null)
{
Question = question;
TrueBranch = trueBranch;
FalseBranch = falseBranch;
Predictions = null;
}
public DecisionTreeId3(DataTable trainingData)
{
ColumnsName = new List <string>();
foreach (DataColumn column in trainingData.Columns)
{
ColumnsName.Add(column.ColumnName);
}
Root = BuildTree(trainingData);
}
public Dictionary <string, float> Query()
{
//Refreshes the data in the features array
GetUpdatedData();
//Printing Tree
Debug.Log("Printing Tree to file: " + Application.persistentDataPath);
XMLManager.PrintTree(this);
DecisionNode leaf = Search(Features.ToList(), rootNode);
if (leaf == null)
{
return(null);
}
return(leaf.Probability);
}
/// <summary>
/// Decide whether to follow the true-branch or the false-branch.
/// Compare the feature / value stored in the node,
/// to the example we're considering.
/// </summary>
private static DecisionNode Classify(Dictionary <string, string> dataRow, DecisionNode node)
{
if (node.IsLeaf())
{
return(node);
}
if (node.Question.IsMatch(dataRow))
{
return(Classify(dataRow, node.TrueBranch));
}
else
{
return(Classify(dataRow, node.FalseBranch));
}
}
/// <summary>
/// World's most elegant tree printing function.
/// </summary>
public static void PrintTree(DecisionNode node, string spacing = "")
{
if (node.IsLeaf())
{
DebugLog(spacing + node, ConsoleColor.Green);
return;
}
DebugLog(spacing + node.Question);
DebugLog(spacing + "--> True:");
PrintTree(node.TrueBranch, spacing + " ");
DebugLog(spacing + "--> False:");
PrintTree(node.FalseBranch, spacing + " ");
}
/// <summary>
/// A nicer way to print the predictions at a leaf.
/// </summary>
private static string PrintLeaf(DecisionNode leaf)
{
var predictions = leaf.Predictions;
var total = predictions.Values.Sum() * 1.0;
var startSymbol = "{";
var finalText = startSymbol;
foreach (var prediction in predictions)
{
if (finalText != startSymbol)
{
finalText += ",";
}
finalText += "\'" + prediction.Key + ": " + (((double)prediction.Value / total) * 100) + "% ";
}
finalText += "}";
return(finalText);
}
private void btnBuildTree_Click(object sender, EventArgs e)
{
List <int> listIndexSet = new List <int>();
for (int i = 0; i < outputs.Length; i++)
{
listIndexSet.Add(i);
}
Dictionary <string, double> probability = Utilities.ProbabilityOfClass(new HashSet <string>(outputs).ToList(), listIndexSet, outputs);
double infoSet = Utilities.Info(probability.Values.ToList());
Bitmap pic = Utilities.CreatePicturePercent(colorForNodes, probability);
treeView1.ImageList = new ImageList();
treeView1.ImageList.Images.Add(pic);
DecisionNode root = new DecisionNode("ALL", "", listIndexSet, infoSet, pictureIndex, probability);
pictureIndex++;
BuildSubTree(root, justificationOfFuzzySet);
treeView1.Nodes.Add(root);
treeView1.Update();
}
/*
* FIXME We do not check each tree for each feature but find it in a first fit method
* This results in inaccurate behaviour but avoids exponential running time...
*/
private DecisionNode Search(List <Feature> features, DecisionNode node)
{
if (features.Count >= 1 && node.DaughterNodes != null && node.DaughterNodes.Count() >= 1)
{
foreach (Feature f in features)
{
//if(node.TestValue.ID == f.ID)
//{
for (int i = 0; i < node.DaughterNodes.Count(); ++i)
{
if (f.TypeOfFeature == node.DaughterNodes[i].FeatureType && f.Actor == node.DaughterNodes[i].Actor)
{
//f.OperatorSign
for (int j = 0; j < node.DaughterNodes.Count(); ++j)
{
if (Math.Abs(float.Parse(node.DaughterNodes[j].incomingEgde) - float.Parse(f.FeatureValue)) < Math.Abs(float.Parse(node.DaughterNodes[i].incomingEgde) - float.Parse(f.FeatureValue)) && i != j)
{
i = j;
}
}
switch (node.DaughterNodes[i].Operator)
{
case ">":
if (float.Parse(f.FeatureValue) > float.Parse(node.DaughterNodes[i].incomingEgde))
{
features.Remove(f);
return(Search(features, node.DaughterNodes[i]));
}
break;
case "<":
if (float.Parse(f.FeatureValue) < float.Parse(node.DaughterNodes[i].incomingEgde))
{
features.Remove(f);
return(Search(features, node.DaughterNodes[i]));
}
break;
case ">=":
if (float.Parse(f.FeatureValue) >= float.Parse(node.DaughterNodes[i].incomingEgde))
{
features.Remove(f);
return(Search(features, node.DaughterNodes[i]));
}
break;
case "<=":
if (float.Parse(f.FeatureValue) <= float.Parse(node.DaughterNodes[i].incomingEgde))
{
features.Remove(f);
return(Search(features, node.DaughterNodes[i]));
}
break;
case "==":
if (float.Parse(f.FeatureValue) == float.Parse(node.DaughterNodes[i].incomingEgde))
{
features.Remove(f);
return(Search(features, node.DaughterNodes[i]));
}
break;
}
}
}
//}
}
}
else
{
return(node);
}
Debug.Log("WARNING - Undefined Behaviour, defaulting to current State");
return(node);
}
private void TreeLearner(TrainingExample[] examples, Feature[] features, DecisionNode decisionNode)
{
if (examples.Length == 0)
{
return;
}
if (features.Length == 0)
{
return;
}
double entropy = Entropy(examples);
List <TrainingExample>[] bestSets = null;
if (entropy <= 0)
{
return;
}
else
{
int exampleCount = examples.Length;
double bestInformationGain = 0;
Feature bestSplitFeature = null;
foreach (Feature f in features)
{
List <TrainingExample>[] sets = SplitByAttribute(examples, f);
double overallEntropy = EntropyOfSets(sets, exampleCount);
double informationGain = entropy - overallEntropy;
if (informationGain > bestInformationGain)
{
bestInformationGain = informationGain;
bestSplitFeature = f;
bestSets = sets;
}
}
decisionNode.TestValue = bestSplitFeature;
List <Feature> lst = features.OfType <Feature>().ToList();
lst.Remove(bestSplitFeature);
Feature[] newFeatures = new Feature[features.Length - 1];
newFeatures = lst.ToArray();
//Instatiate Daugther array of current decisioNode to be equal size of feature's domain
decisionNode.CreateDaughterNodes(bestSplitFeature.FeatureDomain.Count);
for (int i = 0; i < bestSets.Count(); i++)
{
if (bestSets[i].Count == 0)
{
decisionNode.DaughterNodes[i] = new DecisionNode(decisionNode);
decisionNode.DaughterNodes[i].incomingEgde = bestSplitFeature.FeatureDomain[i];
decisionNode.DaughterNodes[i].Action = decisionNode.Action;
decisionNode.DaughterNodes[i].Operator = bestSplitFeature.OperatorSign;
decisionNode.DaughterNodes[i].FeatureType = bestSplitFeature.TypeOfFeature;
decisionNode.DaughterNodes[i].Actor = bestSplitFeature.Actor;
}
else
{
//If there are no trainings examples for the domain value we create a new node with the action.
//Otherwise we create a subset of nodes from each set of trainings examples.
decisionNode.DaughterNodes[i] = new DecisionNode(decisionNode);
decisionNode.DaughterNodes[i].incomingEgde = bestSplitFeature.FeatureDomain[i];
List <State> actions = new List <State>();
foreach (TrainingExample t in bestSets[i])
{
actions.Add(t.Action);
}
State mostCommonAction = (from j in actions
group j by j into grp
orderby grp.Count() descending
select grp.Key).First();
decisionNode.DaughterNodes[i].Action = mostCommonAction;
decisionNode.DaughterNodes[i].FeatureType = bestSplitFeature.TypeOfFeature;
decisionNode.DaughterNodes[i].Probability = GetProbability(actions);
decisionNode.DaughterNodes[i].Operator = bestSplitFeature.OperatorSign;
decisionNode.DaughterNodes[i].Actor = bestSplitFeature.Actor;
TreeLearner(bestSets[i].ToArray(), newFeatures, decisionNode.DaughterNodes[i]);
}
}
}
}
public DecisionNode(DecisionNode parent)
{
this.Parent = parent;
//HACK domain size is only 2
}
private void BuildSubTree(DecisionNode rt, Dictionary <string, Dictionary <string, Dictionary <double, double> > > fuzzySets)
{
List <DecisionNode> ListNodeXk = new List <DecisionNode>();
List <double> GainRatio = new List <double>();
List <double> Gain = new List <double>();
List <double> SplitInfo = new List <double>();
for (int i = 0; i < fuzzySets.Count; i++)
{
GainRatio.Add(0.0);
Gain.Add(0.0);
SplitInfo.Add(0.0);
}
int k = 0;
foreach (var atribut in fuzzySets)
{
int j = Array.FindIndex(inputs, s => s.Equals(atribut.Key));
foreach (var rank in atribut.Value)
{
DecisionNode tmp = new DecisionNode(atribut.Key, rank.Key, new List <int>(), 0.0, -1, new Dictionary <string, double>());
foreach (var i in rt.ListIndexElements)
{
double val = conformityStringToDouble[inputs[j]][data[i, j]];
if (rank.Value[val] >= 0.5)
{
tmp.ListIndexElements.Add(i);
}
}
tmp.ProbabilityClasses = Utilities.ProbabilityOfClass(new HashSet <string>(outputs).ToList(), tmp.ListIndexElements, outputs);
tmp.Entropy = Utilities.Info(tmp.ProbabilityClasses.Values.ToList());
Gain[k] += (Convert.ToDouble(tmp.ListIndexElements.Count) /
Convert.ToDouble(rt.ListIndexElements.Count))
* tmp.Entropy;
double fraq = (double)tmp.ListIndexElements.Count / (double)rt.ListIndexElements.Count;
double lg = Math.Log(fraq) / Math.Log(2);
SplitInfo[k] -= fraq * (Math.Log(fraq) / Math.Log(2));
ListNodeXk.Add(tmp);
}
//вычислить энтропию переменной
Gain[k] = rt.Entropy - Gain[k];
GainRatio[k] = Gain[k] / SplitInfo[k];
k++;
}
//проверить что хотя бы в N дочерних больше чем E элементов
int N = 2;
int E = 2;
int elemMoreThenE = 0;
foreach (var item in ListNodeXk)
{
if (item.ListIndexElements.Count > E)
{
elemMoreThenE++;
}
}
if (elemMoreThenE > N)
{
int Xmax = inputs.ToList().IndexOf(fuzzySets.Keys.ToList()[GainRatio.IndexOf(GainRatio.Max())]);
foreach (var nd in ListNodeXk)
{
if (nd.Atribute == inputs[Xmax])
{
Bitmap pic = Utilities.CreatePicturePercent(colorForNodes, nd.ProbabilityClasses);
treeView1.ImageList.Images.Add(pic);
nd.ImageIndex = pictureIndex;
pictureIndex++;
rt.Nodes.Add(nd);
}
}
Dictionary <string, Dictionary <string, Dictionary <double, double> > > newFuzzySet = new Dictionary <string, Dictionary <string, Dictionary <double, double> > >();
foreach (var item in fuzzySets)
{
if (item.Key != inputs[Xmax])
{
newFuzzySet[item.Key] = item.Value;
}
}
if (newFuzzySet.Count > 0)
{
foreach (DecisionNode curNod in rt.Nodes)
{
if (!curNod.ProbabilityClasses.Values.ToList().Contains(1.0))
{
BuildSubTree(curNod, newFuzzySet);
}
}
}
}
}