/// <summary>
/// Generates the hashcode of this state as if it includes an additional feature.
/// </summary>
/// <param name="withFeature"></param>
/// <returns></returns>
public int GetHashCodeWith(FeatureValuePair withFeature)
{
var newFeatureHashSet = new HashSet <FeatureValuePair>(Features);
newFeatureHashSet.Add(withFeature);
return(GenerateId(newFeatureHashSet).GetHashCode());
}
public override bool Equals(object obj)
{
FeatureValuePair that = (FeatureValuePair)obj;
return(this.Name.Equals(that.Name) &&
this.Value.Equals(that.Value));
}
/// <summary>
/// Creates a Datavector, and adds additional information about feature rewards (or costs) and the classification label.
/// </summary>
/// <param name="headers">The names of the features.</param>
/// <param name="dataobjects">The actual values of each feature.</param>
/// <param name="importance">The relative importance of each feature (-1 to 1).</param>
/// <param name="labelFeatureName">The feature to use as the label. It will be shifted out of the headers and dataobjects and stored as "Label".</param>
public DataVectorTraining(string[] headers, object[] dataobjects, double[] importance, string labelFeatureName)
{
//Check number of headers matches number of data
if ((headers.Length != dataobjects.Length) || (headers.Length != importance.Length))
{
throw new FormatException("Number of headers, importance, and data per line do not match. Ensure there is a header and importance for each value.");
}
//Check label feature is valid
if (!headers.Contains(labelFeatureName))
{
throw new ArgumentException("'labelFeatureName' must exist in the list of headers.");
}
//Build list of features.
Features = new List <FeatureValuePairWithImportance>().Cast <FeatureValuePair>().ToList();
for (int i = 0; i < headers.Length; i++)
{
AddFeature(headers[i], dataobjects[i], importance[i]);
}
//Set label
FeatureValuePair labelFeature = Features.Find(f => f.Name == labelFeatureName);
SetLabel(labelFeature.Name, labelFeature.Value);
//Remove label from the list of features
Features.RemoveAll(p => p.Name == Label.Name);
}
public void RemoveLabel(FeatureValuePair label)
{
lock (processLock)
{
foreach (State state in this.StateSpace.Values)
{
state.RemoveLabel(label);
}
}
}
//Constructors
/// <summary>
/// Creates a new state by combining an existing state and new feature. Queries are updated using the datavector.
/// </summary>
/// <param name="original"></param>
/// <param name="additionalFeature"></param>
/// <param name="dataVector"></param>
public State(State original, FeatureValuePair additionalFeature, DataVectorTraining dataVector) : this(original, dataVector)
{
////Check for disposed and null
//if (additionalFeature == null)
// throw new ArgumentNullException("additionalFeature");
//if (additionalFeature.IsDisposed)
// throw new ArgumentException("Parameter is disposed.", "additionalFeature");
//Add the feature
AddFeature(additionalFeature);
}
public void RemoveStatesWithFeature(FeatureValuePair theFeature)
{
lock (processLock)
{
foreach (State theState in this.StateSpace.Values.ToList())
{
if (theState.Features.Contains(theFeature))
{
this.StateSpace.Remove(theState.GetHashCode());
}
}
}
}
/// <summary>
/// Generates the hashcode of this state as if the specied feature is removed.
/// </summary>
/// <param name="WithoutFeature"></param>
/// <returns></returns>
public int GetHashCodeWithout(FeatureValuePair WithoutFeature)
{
//Copy list of features
var newFeatureHashSet = new HashSet <FeatureValuePair>(Features);
//Remove specified feature
newFeatureHashSet.Remove(WithoutFeature);
//Return zero if no features
if (newFeatureHashSet.Count == 0)
{
return(0);
}
return(GenerateId(newFeatureHashSet).GetHashCode());
}
public void RemoveQueriesWithFeature(FeatureValuePair theFeature)
{
lock (processLock)
{
foreach (State theState in this.StateSpace.Values.ToList())
{
foreach (Query theQuery in theState.Queries.Keys.ToList())
{
if (theQuery.Feature.Equals(theFeature))
{
theState.Queries.Remove(theQuery);
}
}
}
}
}
//Methods
/// <summary>
/// Adds the given feature to the state and removes related queries. This feature is
/// additionaly marked as the "MostRecentFeature" for convienance.
/// </summary>
/// <param name="theFeature"></param>
private void AddFeature(FeatureValuePair theFeature)
{
//Add to list of features
FeatureValuePair fvp = new FeatureValuePair(theFeature.Name, theFeature.Value); //copy to prevent storing derived classes such as FeatureValuePairWithImportance
Features.Add(fvp);
FeatureNames.Add(theFeature.Name);
//Remove queries with same feature name
foreach (var q in Queries.ToList())
{
if (q.Key.Feature.Name == theFeature.Name)
{
Queries.Remove(q.Key);
}
}
}
/// <summary>
/// Updates the percentage probability of each label at this state.
/// </summary>
/// <param name="correctLabel"></param>
public void AdjustLabels(FeatureValuePair correctLabel)
{
//Check impurity. If it is very high, reset the labels.
if (GiniImpurity > 0.99)
{
foreach (var l in Labels.ToList())
{
Labels[l.Key] = 0;
LabelsCount[l.Key] = 0;
}
}
//Reduce label counts occasionally (to prevent going to infinity)
if (LabelsCount.Sum(p => p.Value) > 10000)
{
foreach (FeatureValuePair label in Labels.Select(p => p.Key))
{
LabelsCount[label] /= 10;
}
}
//Add missing label
if (!Labels.ContainsKey(correctLabel))
{
Labels.Add(correctLabel, 0.0);
LabelsCount.Add(correctLabel, 0);
}
//Increase experiences of label
LabelsCount[correctLabel]++;
//Recalculate percentages and gini impurity
double sumCount = LabelsCount.Sum(p => p.Value);
double sumGini = 0;
foreach (var l in Labels.ToList())
{
double labelPercent = LabelsCount[l.Key] / (sumCount); // 0.0 to 1.0
Labels[l.Key] = labelPercent;
sumGini += Math.Pow(labelPercent, 2.0);
}
double maxGini = 1.00000001 - (1.0 / Labels.Count); // 0.000001 prevents division by zero.
GiniImpurity = (1 - sumGini) / maxGini;
}
//Constructors
public Query(FeatureValuePair datavectorFeature, FeatureValuePair label)
{
//Check for nulls
if (datavectorFeature == null || label == null)
{
throw new ArgumentException("Parameters cannot be null.");
}
if (datavectorFeature.Name == null || datavectorFeature.Value == null)
{
throw new ArgumentException("DatavectorFeature's Name and Value parameters cannot be null.");
}
if (label.Name == null || label.Value == null)
{
throw new ArgumentException("Labels's Name and Value parameters cannot be null.");
}
//Save parameters
this.Feature = new FeatureValuePair(datavectorFeature.Name, datavectorFeature.Value); //To prevent additional details being stored by a derived object.
this.Label = new FeatureValuePair(label.Name, label.Value);
}
public void RemoveLabel(FeatureValuePair label)
{
this.Labels.Remove(label);
this.LabelsCount.Remove(label);
}
/// <summary>
/// Selects the best group of queries, then compares them to the appropriate label.
/// If the query's expected reward is better than the label, it returns the query.
/// If the label's expected reward is better, than it returns null, to indicate querying is not the recommended action.
/// </summary>
/// <param name="dataVector"></param>
/// <returns></returns>
public Query GetBestQuery(DataVector dataVector)
{
//Try to add new details
if (dataVector.GetType() == typeof(DataVectorTraining))
{
AddMissingQueriesAndLabels((DataVectorTraining)dataVector);
}
//Get best queries (general)
var bestQueriesGroup = GetAverageGroupQueries();
//Build list of possible queries, that match datavector
var possibleQueries = bestQueriesGroup.Where(q =>
dataVector.Features.Find(f => q.Key.Feature.Equals(f))
!= null
).ToList();
//If no possibilities
if (possibleQueries.Count == 0)
{
return(null);
}
//Result variable
Query bestQueryResult = null; //Default: don't query, because the labels provide the best reward.
#region Find best query, Version 1
//Find best query for each label by expected reward
List <KeyValuePair <Query, double> > bestQueries = new List <KeyValuePair <Query, double> >();
foreach (var labelPair in Labels.ToList())
{
//Get label details
FeatureValuePair theLabel = labelPair.Key;
double theLabelExpectedReward = labelPair.Value;
//Filter list by label
var bestQueriesByLabel = possibleQueries.Where(q => q.Key.Label.Equals(theLabel)).ToList();
if (bestQueriesByLabel.Count == 0)
{
continue;
}
//Get best query details
var bestQueryPair = bestQueriesByLabel.OrderByDescending(p => p.Value).First();
Query bestQuery = bestQueryPair.Key;
double bestQueryExpectedReward = bestQueryPair.Value;
//Is query better than label
if (bestQueryExpectedReward > theLabelExpectedReward)
{
bestQueries.Add(bestQueryPair);
}
}
//Pick final answer
if (bestQueries.Count > 0)
{
bestQueryResult = bestQueries.OrderByDescending(q => q.Value).First().Key;
}
#endregion
#region Find best query, Version 2 -- this may work, and would be faster.
////Find best query pair
//var bestQueryPair2 = possibleQueries.OrderByDescending(p => p.Value).First();
//Query bestQuery2 = bestQueryPair2.Key;
//double bestQuery2ExpectedReward = bestQueryPair2.Value;
////Find label
//var labelPair2 = Labels.ToList().Find(p => p.Key.Equals(bestQuery2.Label));
//Feature theLabel2 = labelPair2.Key;
//double theLabel2ExpectedReward = labelPair2.Value;
////If query has higher expected reward, select it as the option.
//Query bestQueryResult2 = null;
//if (bestQuery2ExpectedReward > theLabel2ExpectedReward)
// bestQueryResult2 = bestQuery2;
#endregion
return(bestQueryResult);
}
public void SetLabel(string featureName, object value)
{
this.Label = new FeatureValuePair(featureName, value);
this.Features.RemoveAll(p => p.Name == featureName);
}
//Methods
public void AddFeature(string featureName, object value)
{
FeatureValuePair fvp = new FeatureValuePair(featureName, value);
Features.Add(fvp);
}