更新1/6/2014:我已经更新了问题，以便尝试求解非线性方程。正如你们中许多人指出的那样，我不需要额外的复杂性(隐藏层，S型函数等)即可解决非线性问题。
另外，我意识到我甚至可以使用除神经网络之外的其他方法来解决非线性问题。我不是在尝试编写效率最高的代码或编写最少的代码。纯粹是为了我更好地学习神经网络。
我创建了自己的反向传播神经网络实现。
经过培训以解决简单的XOR操作时，它工作正常。
但是，现在我想对其进行调整并训练它来解决Y = X * X + B类型公式，但是我没有得到预期的结果。训练后，网络无法计算出正确的答案。神经网络是否适合解决这样的代数方程式？我意识到我的例子很简单，我只是想学习更多有关神经网络及其功能的知识。
我的隐藏层正在使用S型激活函数，而我的输出层正在使用标识函数。
如果您能分析我的代码并指出任何错误，我将不胜感激。
这是我的完整代码(C#.NET):
using System;
using System.Collections.Generic;
using System.Linq;
using System.Text;
namespace NeuralNetwork
{
class Program
{
static void Main(string[] args)
{
Console.WriteLine("Training Network...");
Random r = new Random();
var network = new NeuralNetwork(1, 5, 1);
for (int i = 0; i < 100000; i++)
{
int x = i % 15;
int y = x * x + 10;
network.Train(x);
network.BackPropagate(y);
}
//Below should output 20, but instead outputs garbage
Console.WriteLine("0 * 0 + 10 = " + network.Compute(0)[0]);
//Below should output 110, but instead outputs garbage
Console.WriteLine("10 * 10 + 10 = " + network.Compute(10)[0]);
//Below should output 410, but instead outputs garbage
Console.WriteLine("20 * 20 + 10 = " + network.Compute(20)[0]);
}
}
public class NeuralNetwork
{
public double LearnRate { get; set; }
public double Momentum { get; set; }
public List<Neuron> InputLayer { get; set; }
public List<Neuron> HiddenLayer { get; set; }
public List<Neuron> OutputLayer { get; set; }
static Random random = new Random();
public NeuralNetwork(int inputSize, int hiddenSize, int outputSize)
{
LearnRate = .9;
Momentum = .04;
InputLayer = new List<Neuron>();
HiddenLayer = new List<Neuron>();
OutputLayer = new List<Neuron>();
for (int i = 0; i < inputSize; i++)
InputLayer.Add(new Neuron());
for (int i = 0; i < hiddenSize; i++)
HiddenLayer.Add(new Neuron(InputLayer));
for (int i = 0; i < outputSize; i++)
OutputLayer.Add(new Neuron(HiddenLayer));
}
public void Train(params double[] inputs)
{
int i = 0;
InputLayer.ForEach(a => a.Value = inputs[i++]);
HiddenLayer.ForEach(a => a.CalculateValue());
OutputLayer.ForEach(a => a.CalculateValue());
}
public double[] Compute(params double[] inputs)
{
Train(inputs);
return OutputLayer.Select(a => a.Value).ToArray();
}
public double CalculateError(params double[] targets)
{
int i = 0;
return OutputLayer.Sum(a => Math.Abs(a.CalculateError(targets[i++])));
}
public void BackPropagate(params double[] targets)
{
int i = 0;
OutputLayer.ForEach(a => a.CalculateGradient(targets[i++]));
HiddenLayer.ForEach(a => a.CalculateGradient());
HiddenLayer.ForEach(a => a.UpdateWeights(LearnRate, Momentum));
OutputLayer.ForEach(a => a.UpdateWeights(LearnRate, Momentum));
}
public static double NextRandom()
{
return 2 * random.NextDouble() - 1;
}
public static double SigmoidFunction(double x)
{
if (x < -45.0) return 0.0;
else if (x > 45.0) return 1.0;
return 1.0 / (1.0 + Math.Exp(-x));
}
public static double SigmoidDerivative(double f)
{
return f * (1 - f);
}
public static double HyperTanFunction(double x)
{
if (x < -10.0) return -1.0;
else if (x > 10.0) return 1.0;
else return Math.Tanh(x);
}
public static double HyperTanDerivative(double f)
{
return (1 - f) * (1 + f);
}
public static double IdentityFunction(double x)
{
return x;
}
public static double IdentityDerivative()
{
return 1;
}
}
public class Neuron
{
public bool IsInput { get { return InputSynapses.Count == 0; } }
public bool IsHidden { get { return InputSynapses.Count != 0 && OutputSynapses.Count != 0; } }
public bool IsOutput { get { return OutputSynapses.Count == 0; } }
public List<Synapse> InputSynapses { get; set; }
public List<Synapse> OutputSynapses { get; set; }
public double Bias { get; set; }
public double BiasDelta { get; set; }
public double Gradient { get; set; }
public double Value { get; set; }
public Neuron()
{
InputSynapses = new List<Synapse>();
OutputSynapses = new List<Synapse>();
Bias = NeuralNetwork.NextRandom();
}
public Neuron(List<Neuron> inputNeurons) : this()
{
foreach (var inputNeuron in inputNeurons)
{
var synapse = new Synapse(inputNeuron, this);
inputNeuron.OutputSynapses.Add(synapse);
InputSynapses.Add(synapse);
}
}
public virtual double CalculateValue()
{
var d = InputSynapses.Sum(a => a.Weight * a.InputNeuron.Value) + Bias;
return Value = IsHidden ? NeuralNetwork.SigmoidFunction(d) : NeuralNetwork.IdentityFunction(d);
}
public virtual double CalculateDerivative()
{
var d = Value;
return IsHidden ? NeuralNetwork.SigmoidDerivative(d) : NeuralNetwork.IdentityDerivative();
}
public double CalculateError(double target)
{
return target - Value;
}
public double CalculateGradient(double target)
{
return Gradient = CalculateError(target) * CalculateDerivative();
}
public double CalculateGradient()
{
return Gradient = OutputSynapses.Sum(a => a.OutputNeuron.Gradient * a.Weight) * CalculateDerivative();
}
public void UpdateWeights(double learnRate, double momentum)
{
var prevDelta = BiasDelta;
BiasDelta = learnRate * Gradient; // * 1
Bias += BiasDelta + momentum * prevDelta;
foreach (var s in InputSynapses)
{
prevDelta = s.WeightDelta;
s.WeightDelta = learnRate * Gradient * s.InputNeuron.Value;
s.Weight += s.WeightDelta + momentum * prevDelta;
}
}
}
public class Synapse
{
public Neuron InputNeuron { get; set; }
public Neuron OutputNeuron { get; set; }
public double Weight { get; set; }
public double WeightDelta { get; set; }
public Synapse(Neuron inputNeuron, Neuron outputNeuron)
{
InputNeuron = inputNeuron;
OutputNeuron = outputNeuron;
Weight = NeuralNetwork.NextRandom();
}
}
}
最佳答案
您可以使用Sigmoid作为输出函数，其范围为[0-1]
但是您的目标值是两倍，范围是[0-MAX_INT]，我认为这是您获得NAN的基本原因。
我更新了您的代码，并尝试将[0-1]范围内的值归一化，
我可以得到这样的ouptut，这是我所期望的
我想我接近事实了，我不确定为什么这个答案被否决了
using System;
using System.Collections.Generic;
using System.Linq;
using System.Text;
namespace NeuralNetwork
{
class Program
{
static void Main(string[] args)
{
Console.WriteLine("Training Network...");
Random r = new Random();
var network = new NeuralNetwork(1, 3, 1);
for (int k = 0; k < 60; k++)
{
for (int i = 0; i < 1000; i++)
{
double x = i / 1000.0;// r.Next();
double y = 3 * x;
network.Train(x);
network.BackPropagate(y);
}
double output = network.Compute(0.2)[0];
Console.WriteLine(output);
}
//Below should output 10, but instead outputs either a very large number or NaN
/* double output = network.Compute(3)[0];
Console.WriteLine(output);*/
}
}
public class NeuralNetwork
{
public double LearnRate { get; set; }
public double Momentum { get; set; }
public List<Neuron> InputLayer { get; set; }
public List<Neuron> HiddenLayer { get; set; }
public List<Neuron> OutputLayer { get; set; }
static Random random = new Random();
public NeuralNetwork(int inputSize, int hiddenSize, int outputSize)
{
LearnRate = .2;
Momentum = .04;
InputLayer = new List<Neuron>();
HiddenLayer = new List<Neuron>();
OutputLayer = new List<Neuron>();
for (int i = 0; i < inputSize; i++)
InputLayer.Add(new Neuron());
for (int i = 0; i < hiddenSize; i++)
HiddenLayer.Add(new Neuron(InputLayer));
for (int i = 0; i < outputSize; i++)
OutputLayer.Add(new Neuron(HiddenLayer));
}
public void Train(params double[] inputs)
{
int i = 0;
InputLayer.ForEach(a => a.Value = inputs[i++]);
HiddenLayer.ForEach(a => a.CalculateValue());
OutputLayer.ForEach(a => a.CalculateValue());
}
public double[] Compute(params double[] inputs)
{
Train(inputs);
return OutputLayer.Select(a => a.Value).ToArray();
}
public double CalculateError(params double[] targets)
{
int i = 0;
return OutputLayer.Sum(a => Math.Abs(a.CalculateError(targets[i++])));
}
public void BackPropagate(params double[] targets)
{
int i = 0;
OutputLayer.ForEach(a => a.CalculateGradient(targets[i++]));
HiddenLayer.ForEach(a => a.CalculateGradient());
HiddenLayer.ForEach(a => a.UpdateWeights(LearnRate, Momentum));
OutputLayer.ForEach(a => a.UpdateWeights(LearnRate, Momentum));
}
public static double NextRandom()
{
return 2 * random.NextDouble() - 1;
}
public static double SigmoidFunction(double x)
{
if (x < -45.0)
{
return 0.0;
}
else if (x > 45.0)
{
return 1.0;
}
return 1.0 / (1.0 + Math.Exp(-x));
}
public static double SigmoidDerivative(double f)
{
return f * (1 - f);
}
public static double HyperTanFunction(double x)
{
if (x < -10.0) return -1.0;
else if (x > 10.0) return 1.0;
else return Math.Tanh(x);
}
public static double HyperTanDerivative(double f)
{
return (1 - f) * (1 + f);
}
public static double IdentityFunction(double x)
{
return x;
}
public static double IdentityDerivative()
{
return 1;
}
}
public class Neuron
{
public bool IsInput { get { return InputSynapses.Count == 0; } }
public bool IsHidden { get { return InputSynapses.Count != 0 && OutputSynapses.Count != 0; } }
public bool IsOutput { get { return OutputSynapses.Count == 0; } }
public List<Synapse> InputSynapses { get; set; }
public List<Synapse> OutputSynapses { get; set; }
public double Bias { get; set; }
public double BiasDelta { get; set; }
public double Gradient { get; set; }
public double Value { get; set; }
public Neuron()
{
InputSynapses = new List<Synapse>();
OutputSynapses = new List<Synapse>();
Bias = NeuralNetwork.NextRandom();
}
public Neuron(List<Neuron> inputNeurons)
: this()
{
foreach (var inputNeuron in inputNeurons)
{
var synapse = new Synapse(inputNeuron, this);
inputNeuron.OutputSynapses.Add(synapse);
InputSynapses.Add(synapse);
}
}
public virtual double CalculateValue()
{
var d = InputSynapses.Sum(a => a.Weight * a.InputNeuron.Value);// + Bias;
return Value = IsHidden ? NeuralNetwork.SigmoidFunction(d) : NeuralNetwork.IdentityFunction(d);
}
public virtual double CalculateDerivative()
{
var d = Value;
return IsHidden ? NeuralNetwork.SigmoidDerivative(d) : NeuralNetwork.IdentityDerivative();
}
public double CalculateError(double target)
{
return target - Value;
}
public double CalculateGradient(double target)
{
return Gradient = CalculateError(target) * CalculateDerivative();
}
public double CalculateGradient()
{
return Gradient = OutputSynapses.Sum(a => a.OutputNeuron.Gradient * a.Weight) * CalculateDerivative();
}
public void UpdateWeights(double learnRate, double momentum)
{
var prevDelta = BiasDelta;
BiasDelta = learnRate * Gradient; // * 1
Bias += BiasDelta + momentum * prevDelta;
foreach (var s in InputSynapses)
{
prevDelta = s.WeightDelta;
s.WeightDelta = learnRate * Gradient * s.InputNeuron.Value;
s.Weight += s.WeightDelta; //;+ momentum * prevDelta;
}
}
}
public class Synapse
{
public Neuron InputNeuron { get; set; }
public Neuron OutputNeuron { get; set; }
public double Weight { get; set; }
public double WeightDelta { get; set; }
public Synapse(Neuron inputNeuron, Neuron outputNeuron)
{
InputNeuron = inputNeuron;
OutputNeuron = outputNeuron;
Weight = NeuralNetwork.NextRandom();
}
}
}
关于c# - 使用神经网络求解Y = X * X + b型公式，我们在Stack Overflow上找到一个类似的问题：https://stackoverflow.com/questions/20944178/