/// <summary>
/// 保存したデータを基に学習開始
/// </summary>
public bool Learning(bool isNowTrain)
{
//データが無いなら学習しない
if (situationDatas.Get().Count == 0 && isNowTrain == false)
{
return(false);
}
isTrain = isNowTrain;
if (isNowTrain)
{
isTrain = nn.Train(epochs, learningRate, isNowTrain);
}
else
{
nn.InputData(situationDatas, teachDatas);
//入っているデータを削除
situationDatas = new NumYArray();
teachDatas = new NumYArray();
isTrain = nn.Train(epochs, learningRate, isNowTrain);
}
if (!isTrain)
{
nn.SaveLearningData(charName + "/");
}
return(isTrain);
}
/// <summary>
/// 意思決定
/// </summary>
public int DecideIntention()
{
NumYArray xData = new NumYArray();
xData.Add(situation);
//ニューラルネットワークによる結果表示
NumYArray resultArray = new NumYArray(nn.Predict(xData));
List <List <float> > resultList = resultArray.Get();
//ランダムによる意思決定
float n = Random.Range(0.0f, NumY.Sum(resultArray).Get()[0][0]);
float offset = 0.0f;
for (int i = 0; i < resultList[0].Count; i++)
{
offset += resultList[0][i];
if (n <= offset)
{
intention = i;
break;
}
}
return(this.intention);
}
//予測結果のデバッグ表示
public void DisplayPredict(NumYArray data)
{
string[] name = new string[] { "立ち弱", "立ち強", "しゃがみ弱", "しゃがみ強", "波動", "昇竜",
"ガード", "しゃがみガード", "前進", "後退", "ダッシュ", "ジャンプ", "前ジャンプ", "後ジャンプ" };
for (int i = 0; i < name.Length; i++)
{
Debug.Log(name[i] + " : " + Mathf.Round(data.Get()[0][i] * 100) / 100);
}
}
/// <summary>
/// データを元に結果を予測する
/// </summary>
/// <param name="xData">予測する状況</param>
/// <returns>予測結果</returns>
public NumYArray Predict(NumYArray xData)
{
//順伝播・フォワードプロパゲーション
NumYArray layerZ1 = new NumYArray(NumY.Dot(xData, this.w1) + this.b1);
NumYArray layerA1 = new NumYArray(NumY.Sigmoid(layerZ1));
NumYArray layerZ2 = new NumYArray(NumY.Dot(layerA1, this.w2) + this.b2);
NumYArray layerA2 = new NumYArray(NumY.Sigmoid(layerZ2));
NumYArray returnArray = new NumYArray(layerA2);
return(returnArray);
}
// CSVデータを文字列型２次元配列に変換する
// ファイルパス,変換される配列の値(参照渡し)
public static NumYArray readCSVData(string path)
{
TextAsset csv;
csv = Resources.Load(path) as TextAsset;
StringReader reader = new StringReader(csv.text);
string strStream = reader.ReadToEnd();
// StringSplitOptionを設定(要はカンマとカンマに何もなかったら格納しないことにする)
System.StringSplitOptions option = StringSplitOptions.RemoveEmptyEntries;
// 行に分ける
string[] lines = strStream.Split(new char[] { '\r', '\n' }, option);
// カンマ分けの準備(区分けする文字を設定する)
char[] spliter = new char[1] {
','
};
// 行数設定
int h = lines.Length;
// 列数設定
int w = lines[1].Split(spliter, option).Length;
// 返り値の2次元配列の要素数を設定
List <List <float> > sdata = new List <List <float> >();
// 配列の要素確保
for (int i = 0; i < h - 1; i++)
{
sdata.Add(new List <float>());
for (int j = 0; j < w; j++)
{
sdata[i].Add(0);
}
}
// 行データを切り分けて,2次元配列へ変換する
for (int i = 1; i < h; i++)
{
string[] splitedData = lines[i].Split(spliter, option);
for (int j = 0; j < w; j++)
{
sdata[i - 1][j] = float.Parse(splitedData[j]);
}
}
NumYArray returnArray = new NumYArray(sdata);
return(returnArray);
}
public NumYArray(NumYArray n)
{
array = new List <List <float> >();
List <List <float> > nArray = n.Get();
for (int i = 0; i < nArray.Count; i++)
{
array.Add(new List <float>());
List <float> subList = nArray[i];
for (int j = 0; j < subList.Count; j++)
{
array[i].Add(subList[j]);
}
}
}
/// <summary>
/// 行列の積を求める
/// </summary>
/// <param name="a">自作行列</param>
/// <param name="b">自作行列</param>
/// <returns>計算された自作行列</returns>
public static NumYArray Dot(NumYArray a, NumYArray b)
{
int[] sizeA = a.Size;
int[] sizeB = b.Size;
NumYArray copyA = new NumYArray(a);
NumYArray copyB = new NumYArray(b);
List <List <float> > arrayA = copyA.Get();
List <List <float> > arrayB = copyB.Get();
List <List <float> > array = new List <List <float> >();
//配列の要素確保
for (int i = 0; i < sizeA[0]; i++)
{
array.Add(new List <float>());
for (int j = 0; j < sizeB[1]; j++)
{
array[i].Add(0);
}
}
//行列の内積計算
if (sizeA[1] == sizeB[0])
{
for (int i = 0; i < array.Count; i++)
{
for (int j = 0; j < array[i].Count; j++)
{
float term = 0;
for (int k = 0; k < sizeA[1]; k++)
{
term += arrayA[i][k] * arrayB[k][j];
}
array[i][j] = term;
}
}
}
else
{
//Debug.Log("Error : Argument size does not match.");
}
NumYArray returnArray = new NumYArray(array);
return(returnArray);
}
/// <summary>
/// シグモイド関数による計算
/// </summary>
/// <param name="a">自作配列</param>
/// <returns>計算された自作配列</returns>
public static NumYArray Sigmoid(NumYArray a)
{
NumYArray copyArray = new NumYArray(a);
List <List <float> > array = copyArray.Get();
for (int i = 0; i < array.Count; i++)
{
for (int j = 0; j < array[i].Count; j++)
{
array[i][j] = 1 / (1 + Mathf.Exp(-array[i][j]));
}
}
NumYArray returnArray = new NumYArray(array);
return(returnArray);
}
public NeuralNetwork(int inputUnits, int hiddenUnits, int outpuUnits, string folderName)
{
//学習値のデータがあるならそこから読み込む
if (System.IO.File.Exists(@"Assets/Resources/AIData/" + folderName + "/LearningDataCSVw1.csv"))
{
w1 = NeuralReadCSV.readCSVData("AIData/" + folderName + "/LearningDataCSVw1");
b1 = NeuralReadCSV.readCSVData("AIData/" + folderName + "/LearningDataCSVb1");
w2 = NeuralReadCSV.readCSVData("AIData/" + folderName + "/LearningDataCSVw2");
b2 = NeuralReadCSV.readCSVData("AIData/" + folderName + "/LearningDataCSVb2");
}
else
{
w1 = NumY.RandomArray(inputUnits, hiddenUnits, 1.0f, -1.0f);
b1 = NumY.RandomArray(1, hiddenUnits, 1.0f, -1.0f);
w2 = NumY.RandomArray(hiddenUnits, outpuUnits, 1.0f, -1.0f);
b2 = NumY.RandomArray(1, outpuUnits, 1.0f, -1.0f);
}
}
public static NumYArray RandomArray(int columnNum, int cellNum, float max, float min)
{
List <List <float> > list = new List <List <float> >();
//ランダム配列の生成
for (int i = 0; i < columnNum; i++)
{
list.Add(new List <float>());
for (int j = 0; j < cellNum; j++)
{
list[i].Add(Random.Range(min, max));
}
}
NumYArray returnArray = new NumYArray(list);
return(returnArray);
}
public static NumYArray operator /(NumYArray a, NumYArray b)
{
int[] sizeA = a.Size;
int[] sizeB = b.Size;
NumYArray copyA = new NumYArray(a);
NumYArray copyB = new NumYArray(b);
List <List <float> > arrayA = copyA.Get();
List <List <float> > arrayB = copyB.Get();
//行列の割り算(配列の要素がすべて同じの場合)
if (sizeA[0] == sizeB[0])
{
for (int i = 0; i < sizeA[0]; i++)
{
for (int j = 0; j < sizeA[1]; j++)
{
arrayA[i][j] /= arrayB[i][j];
}
}
}
//(配列の要素の1次元目のみ違う場合)
else if (sizeA[1] == sizeB[1])
{
for (int i = 0; i < sizeA[0]; i++)
{
for (int j = 0; j < sizeA[1]; j++)
{
arrayA[i][j] /= arrayB[0][j];
}
}
}
else
{
Debug.Log("Error : Argument size does not match.");
}
NumYArray array = new NumYArray(arrayA);
return(array);
}
public static NumYArray operator /(NumYArray a, float b)
{
int[] sizeA = a.Size;
NumYArray copyA = new NumYArray(a);
List <List <float> > arrayA = copyA.Get();
for (int i = 0; i < sizeA[0]; i++)
{
for (int j = 0; j < sizeA[1]; j++)
{
arrayA[i][j] /= b;
}
}
NumYArray array = new NumYArray(arrayA);
return(array);
}
public static void SaveCSVData(NumYArray array, string csvName)
{
StreamWriter sw = new StreamWriter(@"Assets/Resources/AIData/" + csvName + ".csv", false, Encoding.GetEncoding("Shift_JIS"));
//ヘッダー書き込み
sw.WriteLine(csvName);
string[] data = new string[array.Size[1]];
//データ書き込み
for (int i = 0; i < array.Size[0]; i++)
{
for (int j = 0; j < array.Size[1]; j++)
{
data[j] = array.Get()[i][j].ToString();
}
string str = string.Join(",", data);
sw.WriteLine(str);
}
sw.Flush();
sw.Close();
}
public static NumYArray operator /(float a, NumYArray b)
{
int[] sizeB = b.Size;
NumYArray copyB = new NumYArray(b);
List <List <float> > arrayB = copyB.Get();
for (int i = 0; i < sizeB[0]; i++)
{
for (int j = 0; j < sizeB[1]; j++)
{
arrayB[i][j] = a / arrayB[i][j];
}
}
NumYArray array = new NumYArray(arrayB);
return(array);
}
// Use this for initialization
void Start()
{
NumYArray xData = new NumYArray(NeuralReadCSV.readCSVData(inputDataName));
NumYArray yData = new NumYArray(NeuralReadCSV.readCSVData(outputDataName));
int epochs = 2000;
float learningRate = 0.1f;
//ニューラルネットワークの生成
NeuralNetwork nn = new NeuralNetwork(xData.Size[1], 8, yData.Size[1], "Aoi");
//時間計測
float startTime = Time.realtimeSinceStartup;
bool a = false;
nn.InputData(xData, yData);
a = nn.Train(epochs, learningRate, a);
//データを基にトレーニング
while (a)
{
a = nn.Train(epochs, learningRate, a);
}
float endTime = Time.realtimeSinceStartup;
//トレーニング時間の表示
Debug.Log(endTime - startTime);
//仮データによる結果の予測
float[][] testList = { new float[] { 0.1f, 0.5f, 0.5f, 0.5f, 0.0f } };
NumYArray testData = new NumYArray(testList);
NumYArray result = new NumYArray(nn.Predict(testData));
//予測結果の表示
DisplayPredict(result);
//学習した重みの保存
nn.SaveLearningData("Aoi/");
}
/// <summary>
/// 初期化
/// </summary>
void Start()
{
charName = gameObject.name.Replace("IntentionObj", "");
string filePath = @"Assets/Resources/AIData/" + charName + "/LearningDataCSVw1.csv";
//csvファイルからデータの読み込み
NumYArray xData = new NumYArray(NeuralReadCSV.readCSVData("AIData/" + inputDataName));
NumYArray yData = new NumYArray(NeuralReadCSV.readCSVData("AIData/" + outputDataName));
//ニューラルネットワークの生成
nn = new NeuralNetwork(xData.Size[1], 8, yData.Size[1], charName);
//学習値のデータがないなら初期学習をする
if (!System.IO.File.Exists(filePath) && !isTrain)
{
nn.InputData(xData, yData);
//データを基にトレーニング
nn.Train(this.epochs, this.learningRate, false);
nn.SaveLearningData(charName + "/");
}
}
/// <summary>
/// 入力されたデータを元にトレーニング
/// </summary>
/// <param name="epochs">トレーニング回数</param>
/// <param name="learningRate">学習率</param>
/// <returns></returns>
public bool Train(int epochs, float learningRate, bool isTrainNow)
{
if (!isTrainNow)
{
NumYArray layerZ1 = new NumYArray();
NumYArray layerA1 = new NumYArray();
NumYArray layerZ2 = new NumYArray();
NumYArray layerA2 = new NumYArray();
NumYArray dlayerZ1 = new NumYArray();
NumYArray dlayerZ2 = new NumYArray();
NumYArray dw1 = new NumYArray();
NumYArray db1 = new NumYArray();
NumYArray dw2 = new NumYArray();
NumYArray db2 = new NumYArray();
totalTrainNum = 0;
maxTrainNum = Mathf.RoundToInt(100 / xData.Get().Count);
if (maxTrainNum <= 0)
{
maxTrainNum = 1;
}
learnSpeed = 1;
}
nowTrainNum = 0;
//規定回数学習させ、他処理のために途中で抜ける
while (nowTrainNum < maxTrainNum * learnSpeed)
{
//許容値を超えているか判定する用出力結果
//Array a = Predict(xData);
int m = xData.Get()[0].Count;
//順伝播・フォワードプロパゲーション
layerZ1 = NumY.Dot(xData, this.w1) + this.b1;
layerA1 = NumY.Sigmoid(layerZ1);
layerZ2 = NumY.Dot(layerA1, this.w2) + this.b2;
layerA2 = NumY.Sigmoid(layerZ2);
//誤差逆伝播法・バックプロパゲーション
dlayerZ2 = (layerA2 - yData) / m;
dw2 = NumY.Dot(layerA1.T, dlayerZ2);
db2 = NumY.Sum(dlayerZ2, 0);
dlayerZ1 = NumY.Dot(dlayerZ2, w2.T) * NumY.SigmoidDerivative(layerZ1);
dw1 = NumY.Dot(xData.T, dlayerZ1);
db1 = NumY.Sum(dlayerZ1, 0);
//パラメータ更新
w2 -= learningRate * dw2;
b2 -= learningRate * db2;
w1 -= learningRate * dw1;
b1 -= learningRate * db1;
nowTrainNum++;
totalTrainNum++;
if (totalTrainNum >= epochs)
{
this.xData = new NumYArray();
this.yData = new NumYArray();
totalTrainNum = 0;
nowTrainNum = 0;
return(false);
}
}
nowTrainNum = 0;
return(true);
}
/// <summary>
/// シグモイド関数の微分
/// </summary>
/// <param name="a">自作配列</param>
/// <returns>微分された自作配列</returns>
public static NumYArray SigmoidDerivative(NumYArray a)
{
NumYArray copyArray = new NumYArray(a);
return(Sigmoid(copyArray) * (1 - Sigmoid(copyArray)));
}
public static NumYArray NumYArray(NumYArray n)
{
return(new NumYArray(n));
}
public void InputData(NumYArray xData, NumYArray yData)
{
this.xData = new NumYArray(xData);
this.yData = new NumYArray(yData);
}
public static NumYArray Sum(NumYArray a, int axis = -1)
{
int[] sizeA = a.Size;
NumYArray copyArray = new NumYArray(a);
List <List <float> > copyList = copyArray.Get();
List <List <float> > list = new List <List <float> >();
//全ての数値を足す
if (axis == -1)
{
float sumNum = 0;
for (int i = 0; i < sizeA[0]; i++)
{
for (int j = 0; j < sizeA[1]; j++)
{
sumNum += copyList[i][j];
}
}
list.Add(new List <float>());
list[0].Add(sumNum);
}
//指定された計算方法による加算
else
{
//列の和を求める
if (axis == 0)
{
list.Add(new List <float>());
for (int i = 0; i < sizeA[1]; i++)
{
float sumNum = 0;
for (int j = 0; j < sizeA[0]; j++)
{
sumNum += copyList[j][i];
}
list[0].Add(sumNum);
}
}
//行の和を求める
else if (axis == 1)
{
for (int i = 0; i < sizeA[0]; i++)
{
list.Add(new List <float>());
float sumNum = 0;
for (int j = 0; j < sizeA[1]; j++)
{
sumNum += copyList[i][j];
}
list[i].Add(sumNum);
}
}
else
{
Debug.Log("Error : Argument size does not match.");
}
}
NumYArray returnArray = new NumYArray(list);
return(returnArray);
}
