### 设计前馈神经网络的BP算法实现 为了设计一个带有BP（反向传播）算法的前馈神经网络，可以使用Python中的`NumPy`库来完成矩阵运算和梯度计算。以下是一个完整的解决方案，支持用户自定义隐含层数、节点数和学习率等超参数。 #### 网络结构与BP算法原理 前馈神经网络是一种典型的多层感知器模型，通过BP算法更新权重和偏置以最小化损失函数。网络由输入层、一个或多个隐含层以及输出层组成。每个层之间的连接权重和偏置通过梯度下降法进行调整[^1]。 ```python import numpy as np class NeuralNetwork: def __init__(self, layers, learning_rate=0.01): """ 初始化神经网络。 :param layers: 网络结构，例如 [input_dim, hidden_dim1, hidden_dim2, output_dim] :param learning_rate: 学习率 """ self.layers = layers self.learning_rate = learning_rate self.weights = [] self.biases = [] # 初始化权重和偏置 for i in range(len(layers) - 1): w = np.random.randn(layers[i], layers[i + 1]) * np.sqrt(2 / layers[i]) b = np.zeros((1, layers[i + 1])) self.weights.append(w) self.biases.append(b) def sigmoid(self, x): """激活函数：Sigmoid""" return 1 / (1 + np.exp(-x)) def sigmoid_derivative(self, x): """Sigmoid导数""" s = self.sigmoid(x) return s * (1 - s) def forward(self, X): """前向传播""" activations = [X] pre_activations = [] for i in range(len(self.weights)): z = np.dot(activations[-1], self.weights[i]) + self.biases[i] a = self.sigmoid(z) pre_activations.append(z) activations.append(a) return activations, pre_activations def compute_loss(self, y_true, y_pred): """均方误差损失""" return np.mean((y_true - y_pred) ** 2) def backward(self, X, y_true, activations, pre_activations): """反向传播""" m = X.shape[0] deltas = [(activations[-1] - y_true) * self.sigmoid_derivative(pre_activations[-1])] # 计算每一层的delta for i in range(len(self.weights) - 1, 0, -1): delta = np.dot(deltas[-1], self.weights[i].T) * self.sigmoid_derivative(pre_activations[i - 1]) deltas.append(delta) deltas.reverse() # 更新权重和偏置 for i in range(len(self.weights)): self.weights[i] -= self.learning_rate * np.dot(activations[i].T, deltas[i]) / m self.biases[i] -= self.learning_rate * np.sum(deltas[i], axis=0, keepdims=True) / m def train(self, X, y_true, epochs): """训练网络""" for epoch in range(epochs): activations, pre_activations = self.forward(X) loss = self.compute_loss(y_true, activations[-1]) if epoch % 100 == 0: print(f"Epoch {epoch}, Loss: {loss}") self.backward(X, y_true, activations, pre_activations) def predict(self, X): """预测""" activations, _ = self.forward(X) return activations[-1] # 示例用法 if __name__ == "__main__": # 定义网络结构 [输入层, 隐含层1, 隐含层2, 输出层] nn = NeuralNetwork(layers=[2, 4, 3, 1], learning_rate=0.1) # 生成一些示例数据 X = np.array([[0, 0], [0, 1], [1, 0], [1, 1]]) y = np.array([[0], [1], [1], [0]]) # 训练网络 nn.train(X, y, epochs=1000) # 测试预测 predictions = nn.predict(X) print("Predictions:") print(predictions) ``` #### 代码说明 - `NeuralNetwork`类实现了前馈神经网络的核心功能，包括初始化、前向传播、损失计算、反向传播和训练。 - 权重和偏置的初始化采用了Xavier初始化方法，确保梯度在传播过程中不会消失或爆炸[^1]。 - 激活函数选择了Sigmoid函数，其导数用于计算梯度。 - 训练过程通过多次迭代逐步降低损失值。 #### 注意事项 - 用户可以通过修改`layers`参数来自定义网络结构，包括隐含层数和每层节点数。 - 学习率`learning_rate`需要根据具体问题调整，过大会导致不稳定，过小会延长收敛时间。