/// <summary>
/// 训练
/// </summary>
/// <param name="netWork"></param>
/// <param name="inputs"></param>
/// <param name="targets"></param>
/// <returns></returns>
public double train(ref BPNN netWork, double[][] inputs, double[][] targets)
{
double result = 0;
for (int i = 0; i < inputs.Length; i++)
{
computeForword(ref netWork, inputs[i]);
double res = backPropagate(ref netWork, targets[i]);
result += res;
}
//Console.WriteLine(j);
return(result / inputs.Length);
}
/// <summary>
/// 前向计算
/// </summary>
/// <param name="netWork">神经网络</param>
/// <param name="inputs">输入值</param>
/// <returns>输出值</returns>
private double[] computeForword(ref BPNN network, double[] inputs)
{
#region 输入层赋值
for (int i = 0; i < inputs.Length; i++)
{
network.i_input[i] = inputs[i];
network.i_output[i] = network.i_input[i].Tanh();
}
network.i_input[network.i_input.Length - 1] = 1;
network.i_output[network.i_output.Length - 1] = network.i_input[network.i_input.Length - 1].Tanh();
#endregion
#region 计算隐藏层
double[] Oij = network.i_output; //上一层输出
for (int i = 0; i < network.h_input.Length; i++)
{
if (i > 0)
{
Oij = network.h_output[i - 1];
}
for (int j = 0; j < network.h_input[i].Length - 1; j++)
{
network.h_input[i][j] = getInputValue(network.h_weights[i], Oij, j);
network.h_output[i][j] = network.h_input[i][j].Tanh();
}
network.h_input[i][network.h_input[i].Length - 1] = 1;
network.h_output[i][network.h_input[i].Length - 1] = (1.0).Tanh();
}
#endregion
#region 计算输出层
Oij = network.h_output[network.h_output.Length - 1];
for (int i = 0; i < network.o_input.Length; i++)
{
network.o_input[i] = getInputValue(network.o_weights, Oij, i);
network.o_output[i] = network.o_input[i].Tanh();
}
#endregion
return(network.o_output);
}
/// <summary>
/// 向后传播，更新权值
/// </summary>
/// <param name="network"></param>
/// <param name="targets"></param>
/// <returns></returns>
private double backPropagate(ref BPNN network, double[] targets)
{
//int row = targets.GetLength(0); //样本数量
//int column = targets.GetLength(1); //样本输出数量
double[] oDelta = new double[targets.Length];
#region 计算输出层Delta
for (int i = 0; i < targets.Length; i++)
{
double error = targets[i] - network.o_output[i];
oDelta[i] = error * (network.o_output[i].DTanh());
}
#endregion
#region 计算隐藏层Delta
double[][] hDeltas = new double[network.h_input.Length][];
for (int i = hDeltas.Length - 1; i >= 0; i--) //从最后一层开始
{
hDeltas[i] = new double[h_input[i].Length];
for (int j = 0; j < hDeltas[i].Length; j++) //j:当前层节点数量，包括了偏置节点
{
double error = 0;
if (i == network.h_input.Length - 1) //是最后一层隐藏层
{
for (int k = 0; k < oDelta.Length; k++)
{
error += oDelta[k] * network.o_weights[j, k];
}
}
else
{
for (int k = 0; k < network.h_output[i + 1].Length - 1; k++) //k:后一层隐藏层节点数量，不包含偏置
{
error += hDeltas[i + 1][k] * network.h_weights[i + 1][j, k];
}
}
hDeltas[i][j] = error * (network.h_output[i][j].DTanh());
}
}
#endregion
#region 更新输出层权值 梯度=输出层Delta*隐藏层输出
double[] lastHiddenOutput = network.h_output[network.h_output.Length - 1]; //最后一层隐藏层输出
for (int j = 0; j < oDelta.Length; j++)
{
for (int i = 0; i < lastHiddenOutput.Length; i++) //i:最后一层隐藏层节点数，包括偏置节点
{
double theta = 0;
theta += oDelta[j] * lastHiddenOutput[i];
network.o_weights[i, j] += theta * network.eta + network.o_lastTheta[i, j] * network.lastEta;
network.o_lastTheta[i, j] = theta;
}
}
#endregion
#region 更新隐藏层权值
for (int i = hDeltas.Length - 1; i >= 0; i--) //从最后一层隐藏层开始更新
{
int lastNeuralCount = i == 0 ? network.i_output.Length : network.h_output[i - 1].Length; //上一层节点数，包含偏置
for (int j = 0; j < hDeltas[i].Length - 1; j++) //j:当前隐藏层节点数，不包含偏置
{
double theta = 0;
for (int k = 0; k < lastNeuralCount; k++)
{
if (i == 0)
{
theta += hDeltas[i][j] * network.i_output[k];
}
else
{
theta += hDeltas[i][j] * network.h_output[i - 1][k];
}
network.h_weights[i][k, j] += network.eta * theta + network.lastEta * network.h_lastTheta[i][k, j];
network.h_lastTheta[i][k, j] = theta;
}
}
}
#endregion
double result = 0.0;
for (int i = 0; i < targets.Length; i++)
{
result += Math.Pow(targets[i] - network.o_output[i], 2);
}
return(result / targets.Length);
}
/// <summary>
/// 预测
/// </summary>
/// <param name="netWork"></param>
/// <param name="inputs"></param>
/// <returns></returns>
public double[] forecast(ref BPNN netWork, double[] inputs)
{
return(computeForword(ref netWork, inputs));
}
