しつこく、chatGPTにXORの学習のC#コードをお願いして(聞くたびに違うプログラムを吐くのが凄いｗ)、やっと動くコードができた。
Unityじゃなくて、Windowsの.NetFrameworkのC#コンパイラーのコンソールアプリでやってみたけど。
XORだけじゃなくて、ANDやORもちゃんと学習出来た
sigmoid dSigmoidを ReLU dReLUに変えても学習出来た
入力を3入力のXORにしても学習出来たので、大丈夫っぽい
これをUnity上で動かせるようにすれば、ゲーム上で、自前でなにか学習できるかも
フレームワークを使うとメモリーとかもかなり食いそうなんで、スマホゲームは自前でやってみてもいいかなと
using System;
namespace NeuralNetwork
{
class Program
{
static void Main(string[] args)
{
// ニューラルネットワークの作成
int[] layerSizes = { 2, 4, 1 };
NeuralNetwork nn = new NeuralNetwork(layerSizes);
double[][] inputs =
{
new double[] { 0, 0 },
new double[] { 0, 1 },
new double[] { 1, 0 },
new double[] { 1, 1 }
};
double[][] targets =
{
new double[] { 0 },
new double[] { 1 },
new double[] { 1 },
new double[] { 0 }
};
// 学習
nn.Train(inputs, targets, 10000, 0.1);
}
}
class NeuralNetwork
{
private int[] layerSizes;
private double[][] activations;
private double[][] outputs;
private double[][][] weights;
private double[][] biases;
private Random rand = new Random();
public NeuralNetwork( int[] layerSizes)
{
this.layerSizes = layerSizes;
activations = new double[layerSizes.Length][];
outputs = new double[layerSizes.Length][];
weights = new double[layerSizes.Length - 1][][];
biases = new double[layerSizes.Length - 1][];
for (int l = 0; l < layerSizes.Length; l++)
{
activations[l] = new double[layerSizes[l]];
outputs[l] = new double[layerSizes[l]];
if( l > 0)
{
weights[l-1] = new double[layerSizes[l]][];
biases[l-1] = new double[layerSizes[l]];
for (int i = 0; i < layerSizes[l]; i++)
{
weights[l - 1][i] = new double[layerSizes[l-1]];
for (int j = 0; j < layerSizes[l - 1]; j++)
{
weights[l - 1][i][j] = rand.NextDouble() * 2 - 1;
}
biases[l - 1][i] = rand.NextDouble() * 2 - 1;
}
}
}
}
public double sigmoid(double x)
{
return 1.0 / ( 1.0 +Math.Exp(-x) );
}
public double dSigmoid(double x)
{
return x * (1.0 - x);
}
public double ReLU(double x)
{
return Math.Max(0,x) ;
}
public double dReLU(double x)
{
return x >0 ? 1 : 0;
}
public double[] FeedForward(double[] input)
{
for (int i = 0; i < layerSizes[0]; i++)
{
activations[0][i] = input[i];
outputs[0][i] = input[i];
}
// 順伝播
for (int l = 1; l < layerSizes.Length; l++)
{
for (int i = 0; i < layerSizes[l]; i++)
{
double sum = biases[l - 1][i];
for (int j = 0; j < layerSizes[l - 1]; j++)
{
sum += weights[l - 1][i][j] * outputs[l - 1][j];
}
activations[l][i] = sum;
outputs[l][i] = sigmoid(sum);
}
}
// 出力を返す
return outputs[layerSizes.Length - 1];
}
private void BackPropagation(double[] target, double learningRate)
{
// 出力層の誤差を計算
for (int i = 0; i < layerSizes[layerSizes.Length - 1]; i++)
{
double output = outputs[layerSizes.Length - 1][i];
double error = target[i] - output;
double derivative = dSigmoid( output );
activations[layerSizes.Length - 1][i] = error * derivative;
}
// 逆伝播
for (int l = layerSizes.Length - 2; l >= 1; l--)
{
for (int i = 0; i < layerSizes[l]; i++)
{
double output = outputs[l][i];
double derivative = dSigmoid( output );
double sum = 0;
for (int j = 0; j < layerSizes[l + 1]; j++)
{
sum += activations[l + 1][j] * weights[l][j][i];
}
activations[l][i] = sum * derivative;
}
}
// 重みとバイアスの更新
for (int l = 0; l < layerSizes.Length - 1; l++)
{
for (int i = 0; i < layerSizes[l + 1]; i++)
{
for (int j = 0; j < layerSizes[l]; j++)
{
double output = outputs[l][j];
double delta = learningRate * activations[l + 1][i] * output;
weights[l][i][j] += delta;
}
biases[l][i] += learningRate * activations[l + 1][i];
}
}
}
// 学習
public void Train(double[][] inputs, double[][] targets, int iterations, double learningRate)
{
for (int i = 0; i < iterations; i++)
{
for (int j = 0; j < inputs.Length; j++)
{
double[] output = FeedForward(inputs[j]);
BackPropagation(targets[j], learningRate);
}
// テスト
if( (i%1000) == 0)
{
String str = "";
for (int j = 0; j < 4; j++)
{
str += inputs[j][0]+"x"+inputs[j][1]+" = "+FeedForward(inputs[j])[0]+", ";
}
Console.WriteLine( "#"+i+" " + str);
}
}
}
}
}
