完整代码
the full code of the neuron. am trying to code a radial basis neuron network but i can figure out where and getting an NaN error
import java.util.ArrayList; import java.util.List; import java.util.Random;
public class Neurons {
static Random rand =new Random();
instance variables
private float m_neuronOutput;
private float alpha=0.4f;
private float m_gradient;
private int neuronIndex;
public float eta;
public float sigma=22.5f;
public List<Connection> neuronWeights= new ArrayList<Connection>();
//public List<Center> neuronCenter= new ArrayList<Center>();
NetWork net=new NetWork();
public Neurons(){
// sigma=22.5f;
eta=0.5f;
}
//neuron constructor
public Neurons(int outputNum, int m_neuronIndex){
neuronIndex=m_neuronIndex;
add connection weights
for (int connect = 0; connect < outputNum; connect++) {
neuronWeights.add(new Connection());
}
}
public void SetOutputValue(float value){
m_neuronOutput = value;
}
public float GetOutputValue(){
return m_neuronOutput;
}
## feed the layer with the input ##
提供图层
public void FeedForward(Layer previousLayer, int n){
float sum = 0;
//weightsum the inputs of the last layer, including bias
//System.out.println(n);
m_neuronOutput= previousLayer.neuron.get(n).GetOutputValue () ;
//System.out.println(m_neuronOutput);
}
设置隐藏图层的值
public void FeedHidden(Layer previousLayer, int j ){
// for(int i=0;i<center.length;i++)
// System.out.println(center[i]);
float distance= 0;
//System.out.println(n+" "+hiddenLayer.neuron.get(n).neuronCenter.get(n).center);
//weightsum the inputs of the last layer, including bias
for (int n = 0; n <previousLayer.neuron.size(); n++) {
distance += (float) Math.pow(previousLayer.neuron.get(n).GetOutputValue ()-previousLayer.neuronCenter.get(j).center,2);
//System.out.println(previousLayer.neuronCenter.get(j).center);
}
## where am calling phi(Gaussian function) ## m_neuronOutput = phi (distance);
//System.out.println(m_neuronOutput);
}
private float phi( float distance){
//System.out.println(sigma);
return (float) Math.exp(- distance/(2*Math.pow(sigma, 2)));
}
计算输出
public void setOutput(Layer hidden){
float output=0;
for(int i=0;i<hidden.neuron.size();i++){
output+=hidden.neuron.get(i).GetOutputValue()*hidden.neuron.get(i).neuronWeights.get(neuronIndex).weight;
// System.out.println(hidden.neuron.get(i).GetOutputValue());
// System.out.println(hidden.neuron.get(i).neuronWeights.get(neuronIndex).weight);
}
m_neuronOutput=output;
//System.out.println(hidden.neuron.get(i).neuronWeights.get(neuronIndex).weight);
}
输出图层的渐变
public void CalcOutputGradients( float targetValue,float output){
// System.out.println(output);
float delta = targetValue -output;
m_gradient = delta;
//System.out.println( output);
}
public void UpdateInputWeights(Layer previousLayer,float eta){
//System.out.println(phi);
the weights to be updated are in the connection container
in the neurons in the preceding layer
int numNeuronas = previousLayer.neuron.size();
for (int n = 0; n < numNeuronas; n++) {//for every previous layer neuron
Neurons neu = previousLayer.neuron.get(n);//getting the previous layer neuron
//System.out.println(neu.neuronCenter.get(neuronIndex).center );
float phi=previousLayer.neuron.get(n).GetOutputValue ();
//double oldDeltaW = neu.neuronWeights.get(neuronIndex).deltaweight;
double deltaWeight=
//Individual input magnified by the gradient and trainrate
eta *phi*m_gradient;
also add momentum = a fraction of the previos delta weight
+oldDeltaW;
neu.neuronWeights.get(neuronIndex).deltaweight = deltaWeight;
neu.neuronWeights.get(neuronIndex).weight+= deltaWeight;
//System.out.println(m_gradient +" delta");
}
}
## my method to update the center but yet to call it in my network work class ##
public void updateCenter( Layer previousLayer,float[] input, float eta){
int numNeuronas = previousLayer.neuron.size();
for (int n = 0; n < numNeuronas-1; n++) {//for every previous layer neuron
Neurons neu = previousLayer.neuron.get(n);//getting the previous layer neuro
Center neu1 = previousLayer.neuronCenter.get(n);
float phi=previousLayer.neuron.get(n).GetOutputValue ();
double oldcenter = neu1.center;
double weight= neu.neuronWeights.get(neuronIndex).weight;
double deltaC=
//Individual input magnified by the gradient and trainrate
eta *phi*((m_gradient*weight)/Math.pow(sigma, 2))*(input[n]-oldcenter);
//also add momentum = a fraction of the previos delta weight
//+oldcenter;
neu1.center+= deltaC;
//System.out.println(gredient+" delta");
}
}
public void updateSigma(Layer previousLayer, float[] center, float[] input, float eta){
float distance= 0;
//weightsum the inputs of the last layer, including bias
for (int n = 0; n < center.length; n++) {
distance+= Math.pow(previousLayer.neuron.get(n).GetOutputValue () -center[n],2);
//System.out.println(distance);
}
float phi=(float) phi(distance);
float sigma1=0;
for(int i=0;i<center.length;i++){
double w =previousLayer.neuron.get(i).neuronWeights.get(neuronIndex).weight;
sigma1=(float) (sigma+(eta*((m_gradient*w)/Math.pow(sigma, 3))*phi*Math.pow(input[i]-center[i],2)));
}
sigma=sigma1;
}
get the current learning rate
public float updateLearningRate(float epoch){
//search the converge algorithm
return (float) (eta/((epoch)/500));
}
set the learning rate
public void setLearningRate(float rate){
eta=rate;
}
update the learining rate
public float getLearningRate(){
return eta;
}
}
寻求帮助。感谢
答案 0 :(得分：2)
return (float) Math.exp(- distance/(2*Math.pow(sigma, 2)));
在上一行中，当sigma和distance为0时，您执行Math.exp(-0/0) NaN。
NaN代表Not a Number，例如当你除以0或取一个负数的平方根时。
NaN由 IEE 754 定义，代表双倍，实际上并非数值。
作为旁注，我相信您会阅读，请始终将堆栈跟踪添加到问题中。始终添加有关问题的所有重要信息。