/// <summary>
/// Searches for the best rendition given the variables and constraints.
/// </summary>
private Ptype.Mutable Search(State state)
{
//Removes any paths that conflict with the given constraints
if (!Propagate(state.Parts.Values, state.Constraints))
{
return(state.BestRendition);
}
//Computes the cost of the full or partial assignment
double cost = _costFunction.Cost(state.Parts, state.PositiveExamples, state.NegativeExamples, state.Cache);
if (cost >= state.BestCost)
{
return(state.BestRendition);
}
//If the assignment is complete, then we know it's
//the best cost so far, so assign it as the best rendition
if (CompleteAssignment(state.Parts.Values))
{
Ptype.Mutable rendition = GetPrototype(state);
if (!MatchesAnyNegatives(rendition, state.NegativeExamples))
{
state.BestCost = cost;
state.BestRendition = rendition;
}
return(state.BestRendition);
}
//Select the next variable to assign
Part partToAssign = _orderSelecter.SelectNextPartToAssign(state.Parts, state.PositiveExamples, state.NegativeExamples);
//Copy the variable's values so we can iterate through them
ArrayList partValues = new ArrayList(partToAssign.CurrentValidValues);
//Assign each value to the variable and continue searching
foreach (object value in partValues)
{
partToAssign.AssignValue(value);
Search(state);
UndoPropogate(state.Parts.Values);
}
//Give the variable all of it's values back
partToAssign.UnassignValueAndRestoreDomain();
//Put all of the possible search paths back
UndoPropogate(state.Parts.Values);
return(state.BestRendition);
}
/// <summary>
/// Calculates the cost of the specified prediction.
/// </summary>
/// <param name="prediction">The prediction.</param>
/// <param name="oneHot">The one hot value.</param>
/// <returns>The cost.</returns>
public virtual double CalculateError(double[] prediction, int oneHot)
{
double error = _costFunction.Cost(prediction, oneHot.OneHot(prediction.Length));
return(error);
}