/**
* 代入函数
* @param feature 特征X（维度由构造时的index指定）
* @param label Y
* @return
*/
public int Apply(MLFeature <FeatureT> feature, LabelT label)
{
if (feature[index].Equals(value) && label.Equals(this.label))
{
return(1);
}
return(0);
}
////
/// <summary>
/// 求delta_i
/// </summary>
/// <param name="empirical_e">@param empirical_e fi的期望</param>
/// <param name="fi">@param fi fi的下标</param>
/// <returns></returns>
//private double iis_solve_delta(double empirical_e, int fi)
private double iis_solve_delta(double empirical_e, string fi)
{
double delta = 0;
double f_newton, df_newton;
double[,] p_yx = calc_prob_y_given_x();
int iters = 0;
while (iters < 50) // 牛顿法
{
f_newton = df_newton = 0;
for (int i = 0; i < instances.Count; i++)
{
MLInstance <int, int> instance = instances[i];
MLFeature <int> tfeature = instance.Feature;
Feature feature = new Feature(tfeature);
int index = features.IndexOf(feature);
if (index == -1)
{
index = feature.getIndex(instance.Feature.Count);
if (string.Join("", features[index]) == string.Join("", feature))
{
features[index] = feature;
}
}
for (int y = minY; y <= maxY; y++)
{
int f_sharp = apply_f_sharp(feature, y);
double prod = p_yx[index, y] * functions[fi].Apply(feature, y) * Math.Exp(delta * f_sharp);
f_newton += prod;
df_newton += prod * f_sharp;
}
}
f_newton = empirical_e - f_newton / N; // g
df_newton = -df_newton / N; // g的导数
if (Math.Abs(f_newton) < 0.0000001)
{
return(delta);
}
double ratio = f_newton / df_newton;
delta -= ratio;
if (Math.Abs(ratio) < EPSILON)
{
return(delta);
}
iters++;
}
return(double.NaN); //如果不收敛，返回NaN
throw new Exception("IIS did not converge"); // w_i不收敛
}
/// <summary>
/// 计算经验期望
/// </summary>
private void calc_empirical_expects()
{
foreach (MLInstance <int, int> instance in instances)
{
int y = instance.Label;
MLFeature <int> feature = instance.Feature;
//for (int i = 0; i < functions.Count; i++)
int i = 0;
foreach (string key in functions.Keys)
{
empirical_expects[i] += functions[key].Apply(feature, y);
i++;
}
}
for (int i = 0; i < functions.Count; i++)
{
empirical_expects[i] /= 1.0 * N;
}
//if (DEBUG) System.out.println(Arrays.toString(empirical_expects));
}
public MLInstance(MLFeature <FeatureT> ft)
{
Feature = ft;
}
public MLInstance(LabelT label, List <FeatureT> flist)
{
Feature = new MLFeature <FeatureT>();
Label = label;
}
public MLInstance(List <FeatureT> flist)
{
Feature = new MLFeature <FeatureT>(flist);
}
