/// <summary>
/// 后向传播算法
/// </summary>
/// <param name="targets"></param>
/// <param name="N"></param>
/// <param name="M"></param>
/// <returns></returns>
public double BackPropagate(double[] targets, double N, double M)
{
//计算输出层 deltas
double[] output_deltas = new double[this._no];
for (int k = 0; k < this._no; k++)
{
double oerror = targets[k] - this._ao[k];
output_deltas[k] = oerror * MathFunction.DerivativeSigmoid(this._ao[k]);
}
//更新输出层权值
for (int j = 0; j < this._nh; j++)
{
for (int k = 0; k < this._no; k++)
{
double change = output_deltas[k] * this._ah[j];
this._wo[j][k] += N * change + M * this._co[j][k];
this._co[j][k] = change;
}
}
//计算隐藏层 deltas
double[] hidden_deltas = new double[this._nh];
for (int j = 0; j < this._nh; j++)
{
double herror = 0.0;
for (int k = 0; k < this._no; k++)
{
herror += output_deltas[k] * this._wo[j][k];
}
hidden_deltas[j] = herror * MathFunction.DerivativeSigmoid(this._ah[j]);
}
//更新输入层权值
for (int i = 0; i < this._ni; i++)
{
for (int j = 0; j < this._nh; j++)
{
double change = hidden_deltas[j] * this._ai[i];
this._wi[i][j] += N * change + M * this._ci[i][j];
this._ci[i][j] = change;
}
}
//计算误差平方和
double error = 0.0;
for (int k = 0; k < targets.Length; k++)
{
error = 0.5 * Math.Pow((targets[k] - this._ao[k]), 2);
}
return(error);
}
/// <summary>
/// 构造神经网络
/// </summary>
/// <param name="ni">输入单元数量</param>
/// <param name="nh">隐藏单元数量</param>
/// <param name="no">输出单元数量</param>
public void Init(int ni, int nh, int no)
{
//各层单元数量
this._ni = ni + 1;
this._nh = nh;
this._no = no;
//激活值（输出值）
this._ai = new double[this._ni];
this._ah = new double[this._nh];
this._ao = new double[this._no];
//创建权重矩阵
this._wi = MathFunction.MakeMatrix(this._ni, this._nh);
this._wo = MathFunction.MakeMatrix(this._nh, this._no);
//将权重矩阵随机化
MathFunction.RandomizeMatrix(this._wi);
MathFunction.RandomizeMatrix(this._wo, -1, 1);
//权重矩阵的上次梯度
this._ci = MathFunction.MakeMatrix(this._ni, this._nh);
this._co = MathFunction.MakeMatrix(this._nh, this._no);
}
