麻雀算法SSA优化LSTM超参数

前言

  1. LSTM 航空乘客预测单步预测的两种情况。 简单运用LSTM 模型进行预测分析。
  2. 加入注意力机制的LSTM 对航空乘客预测采用了目前市面上比较流行的注意力机制,将两者进行结合预测。
  3. 多层 LSTM 对航空乘客预测 简单运用多层的LSTM 模型进行预测分析。
  4. 双向LSTM 对航空乘客预测双向LSTM网络对其进行预测。
  5. MLP多层感知器 对航空乘客预测简化版 使用MLP 对航空乘客预测
  6. CNN + LSTM 航空乘客预测采用的CNN + LSTM网络对其进行预测。
  7. ConvLSTM 航空乘客预测采用ConvLSTM 航空乘客预测
  8. LSTM的输入格式和输出个数说明 中对单步和多步的输入输出格式进行了解释
  9. LSTM 单变量多步预测航空乘客简单版
  10. LSTM 单变量多步预测航空乘客复杂版
  11. LSTM 多变量单步预测空气质量(1—》1) 用LSTM 前一个数据点的多变量预测下一个时间点的空气质量
  12. LSTM 多变量单步预测空气质量(3 —》1) 用LSTM 前三个数据点的多变量预测下一个时间点的空气质量

本文主要是采用麻雀算法SSA优化LSTM超参数

程序

麻雀搜索算法是2020提出的一种新的优化算法,在此不对具体原理进行分析,针对代码实操.

SSA

麻雀算法代码简介

class SSA():
    def __init__(self, func, n_dim=None, pop_size=20, max_iter=50, lb=-512, ub=512, verbose=False):
        self.func = func
        self.n_dim = n_dim  # dimension of particles, which is the number of variables of func
        self.pop = pop_size  # number of particles
        P_percent = 0.2  # # 生产者的人口规模占总人口规模的20%
        D_percent = 0.1  # 预警者的人口规模占总人口规模的10%
        self.pNum = round(self.pop * P_percent)  # 生产者的人口规模占总人口规模的20%
        self.warn = round(self.pop * D_percent)  # 预警者的人口规模占总人口规模的10%

        self.max_iter = max_iter  # max iter
        self.verbose = verbose  # print the result of each iter or not

        self.lb, self.ub = np.array(lb) * np.ones(self.n_dim), np.array(ub) * np.ones(self.n_dim)
        assert self.n_dim == len(self.lb) == len(self.ub), 'dim == len(lb) == len(ub) is not True'
        assert np.all(self.ub > self.lb), 'upper-bound must be greater than lower-bound'

        self.X = np.random.uniform(low=self.lb, high=self.ub, size=(self.pop, self.n_dim))

        self.Y = [self.func(self.X[i]) for i in range(len(self.X))]  # y = f(x) for all particles
        self.pbest_x = self.X.copy()  # personal best location of every particle in history
        self.pbest_y = [np.inf for i in range(self.pop)]  # best image of every particle in history
        self.gbest_x = self.pbest_x.mean(axis=0).reshape(1, -1)  # global best location for all particles
        self.gbest_y = np.inf  # global best y for all particles
        self.gbest_y_hist = []  # gbest_y of every iteration
        self.update_pbest()
        self.update_gbest()
        #
        # record verbose values
        self.record_mode = False
        self.record_value = {'X': [], 'V': [], 'Y': []}
        self.best_x, self.best_y = self.gbest_x, self.gbest_y  # history reasons, will be deprecated
        self.idx_max = 0
        self.x_max = self.X[self.idx_max, :]
        self.y_max = self.Y[self.idx_max]

    def cal_y(self, start, end):
        # calculate y for every x in X
        for i in range(start, end):
            self.Y[i] = self.func(self.X[i])
        # return self.Y

    def update_pbest(self):
        '''
        personal best
        '''
        for i in range(len(self.Y)):
            if self.pbest_y[i] > self.Y[i]:
                self.pbest_x[i] = self.X[i]
                self.pbest_y[i] = self.Y[i]

    def update_gbest(self):
        idx_min = self.pbest_y.index(min(self.pbest_y))
        if self.gbest_y > self.pbest_y[idx_min]:
            self.gbest_x = self.X[idx_min, :].copy()
            self.gbest_y = self.pbest_y[idx_min]

    def find_worst(self):
        self.idx_max = self.Y.index(max(self.Y))
        self.x_max = self.X[self.idx_max, :]
        self.y_max = self.Y[self.idx_max]

    def update_finder(self):
        r2 = np.random.rand(1)  # 预警值
        self.idx = sorted(enumerate(self.Y), key=lambda x: x[1])
        self.idx = [self.idx[i][0] for i in range(len(self.idx))]
        # 这一部位为发现者(探索者)的位置更新
        if r2 < 0.8:  # 预警值较小,说明没有捕食者出现
            for i in range(self.pNum):
                r1 = np.random.rand(1)
                self.X[self.idx[i], :] = self.X[self.idx[i], :] * np.exp(-(i) / (r1 * self.max_iter))  # 对自变量做一个随机变换
                self.X = np.clip(self.X, self.lb, self.ub)  # 对超过边界的变量进行去除
                # X[idx[i], :] = Bounds(X[idx[i], :], lb, ub)  # 对超过边界的变量进行去除
                # fit[sortIndex[0, i], 0] = func(X[sortIndex[0, i], :])  # 算新的适应度值
        elif r2 >= 0.8:  # 预警值较大,说明有捕食者出现威胁到了种群的安全,需要去其它地方觅食
            for i in range(self.pNum):
                Q = np.random.rand(1)  # 也可以替换成  np.random.normal(loc=0, scale=1.0, size=1)
                self.X[self.idx[i], :] = self.X[self.idx[i], :] + Q * np.ones(
                    (1, self.n_dim))  # Q是服从正态分布的随机数。L表示一个1×d的矩阵
                self.X = np.clip(self.X, self.lb, self.ub)  # 对超过边界的变量进行去除
                # X[idx[i], :] = Bounds(X[sortIndex[0, i], :], lb, ub)
                # fit[sortIndex[0, i], 0] = func(X[sortIndex[0, i], :])
        self.cal_y(0, self.pNum)

    def update_follower(self):
        #  这一部位为加入者(追随者)的位置更新
        for ii in range(self.pop - self.pNum):
            i = ii + self.pNum
            A = np.floor(np.random.rand(1, self.n_dim) * 2) * 2 - 1
            best_idx = self.Y[0:self.pNum].index(min(self.Y[0:self.pNum]))
            bestXX = self.X[best_idx, :]
            if i > self.pop / 2:
                Q = np.random.rand(1)
                self.X[self.idx[i], :] = Q * np.exp((self.x_max - self.X[self.idx[i], :]) / np.square(i))
            else:
                self.X[self.idx[i], :] = bestXX + np.dot(np.abs(self.X[self.idx[i], :] - bestXX),
                                                         1 / (A.T * np.dot(A, A.T))) * np.ones((1, self.n_dim))
        self.X = np.clip(self.X, self.lb, self.ub)  # 对超过边界的变量进行去除
        # X[self.idx[i],:] = Bounds(X[self.idx[i],lb,ub)
        # fit[self.idx[i],0] = func(X[self.idx[i], :])
        self.cal_y(self.pNum, self.pop)

    def detect(self):
        arrc = np.arange(self.pop)
        c = np.random.permutation(arrc)  # 随机排列序列
        b = [self.idx[i] for i in c[0: self.warn]]
        e = 10e-10
        for j in range(len(b)):
            if self.Y[b[j]] > self.gbest_y:
                self.X[b[j], :] = self.gbest_y + np.random.rand(1, self.n_dim) * np.abs(self.X[b[j], :] - self.gbest_y)
            else:
                self.X[b[j], :] = self.X[b[j], :] + (2 * np.random.rand(1) - 1) * np.abs(
                    self.X[b[j], :] - self.x_max) / (self.func(self.X[b[j]]) - self.y_max + e)
            # X[sortIndex[0, b[j]], :] = Bounds(X[sortIndex[0, b[j]], :], lb, ub)
            # fit[sortIndex[0, b[j]], 0] = func(X[sortIndex[0, b[j]]])
            self.X = np.clip(self.X, self.lb, self.ub)  # 对超过边界的变量进行去除
            self.Y[b[j]] = self.func(self.X[b[j]])

    def run(self, max_iter=None):
        self.max_iter = max_iter or self.max_iter
        for iter_num in range(self.max_iter):
            self.update_finder()  # 更新发现者位置
            self.find_worst()  # 取出最大的适应度值和最差适应度的X
            self.update_follower()  # 更新跟随着位置
            self.update_pbest()
            self.update_gbest()
            self.detect()
            self.update_pbest()
            self.update_gbest()
            self.gbest_y_hist.append(self.gbest_y)
        return self.best_x, self.best_y

LSTM

def build_model(neurons1, neurons2, dropout):
    X_train, y_train, X_test, y_test = process_data()
    # X_train, y_train = create_dataset(X_train, y_train, steps)
    # X_test, y_test = create_dataset(X_test, y_test, steps)
    nb_features = X_train.shape[2]
    input1 = X_train.shape[1]
    model1 = Sequential()
    model1.add(LSTM(
        input_shape=(input1, nb_features),
        units=neurons1,
        return_sequences=True))
    model1.add(Dropout(dropout))

    model1.add(LSTM(
        units=neurons2,
        return_sequences=False))
    model1.add(Dropout(dropout))

    model1.add(Dense(units=1))
    model1.add(Activation("linear"))
    model1.compile(loss='mse', optimizer='Adam', metrics='mae')
    return model1, X_train, y_train, X_test, y_test

优化超参数

if __name__ == '__main__':
    '''
    神经网络第一层神经元个数
    神经网络第二层神经元个数
    dropout比率
    batch_size
    '''
    neurons1 = 64
    neurons2 = 64
    dropout = 0.01
    batch_size = 32
    model, X_train, y_train, X_test, y_test = build_model(neurons1, neurons2, dropout)
    history1 = model.fit(X_train, y_train, epochs=150, batch_size=batch_size, validation_split=0.2, verbose=1,
                         callbacks=[EarlyStopping(monitor='val_loss', patience=9, restore_best_weights=True)])
    # 测试集预测
    y_score = model.predict(X_test)
    # 反归一化
    y_score = scaler.inverse_transform(y_score.reshape(-1, 1))
    y_test = scaler.inverse_transform(y_test.reshape(-1, 1))

    print("==========evaluation==============\n")
    from sklearn.metrics import mean_squared_error
    from sklearn.metrics import mean_absolute_error #平方绝对误差
    import math

    MAE = mean_absolute_error(y_test, y_score)
    print('MAE: %.4f ' % MAE)
    RMSE = math.sqrt(mean_squared_error(y_test, y_score))
    print('RMSE: %.4f ' % (RMSE))
    

总结

  1. SSA在一定范围内可以优化LSTM 的超参数,对算力要求有点大
  2. SSA优化算法有一定的局限性,如何利用其优势至关重要
  3. LSTM的超参数可以部分优化,能够节约时间和节省算力资源

备注:
需要源代码和数据集,或者想要沟通交流,请私聊,谢谢.

你可能感兴趣的:(机器学习算法研究,时间序列预测,算法,lstm,深度学习)