WhiteHacker 2019-06-22
人工神经网络的灵感来自我们的大脑。遗传算法受到进化的启发。本文提出了一种新型的辅助训练的神经网络:遗传神经网络。这些神经网络具有适应度等特性,并使用遗传算法训练随机生成的权重。遗传优化发生在任何形式的反向传播之前,以给梯度下降提供一个更好的起点。
序列神经网络接受一个输入矩阵,在模型外部与一个真实输出值的向量配对。然后通过遍历每一层,通过权重和激活函数来变换矩阵。
这是一个序列神经网络,具有一个输入矩阵,两个隐藏层,一个输出层,三个权重矩阵和一种激活函数。
最初的预测很可能是不准确的,所以为了训练一个序列神经网络做出更好的预测,我们把它看作一个复合函数。
创建一个损失函数,输入矩阵和真实输出向量(X和y)保持不变。
现在所有的东西都是关于函数的,并且有一个明确的目标(最小化损失),我们得到一个多变量微积分的优化问题。
随着模型显示出越来越多的复杂性,梯度下降的计算成本可能变得非常昂贵。遗传神经网络提供了一个可供选择的初始训练过程,以提供一个更好的起点,在反向传播过程中允许更少的epochs。
在遗传神经网络中,网络被视为具有fields和适应度的计算对象。这些fields被认为是在反向传播之前通过遗传算法优化的基因。这使得梯度下降具有更好的起始位置,并且允许更少的训练时间,并具有更高的模型测试准确度。考虑以下遗传神经网络,其中权重被视为计算对象中的fields。
这些fields是相对于遗传神经网络的每个实例的基因。就像序列神经网络一样,它可以表示为复合函数。
然而,在使用微积分之前,我们将使用遗传算法采取进化方法来优化权重。
在自然界中,染色体交叉看起来是这样的…
如果我们把染色体简化成块…
这与遗传算法用于改变权重矩阵的逻辑相同。这个想法将是创建一个初始种群的n个遗传神经网络,经过正向传播计算出一个适应度得分,最后选择最适合的个体来创建孩子。这个过程将重复,直到找到最优的初始权值进行反向传播。
首先,我们必须建立遗传神经网络。我们使用的是具有四个输入节点,两个隐藏层和一个输出层的训练模型(以匹配上图),这可以扩展到任何类型的神经网络。
import pandas as pd import numpy as np import random from sklearn.model_selection import train_test_split from sklearn.metrics import accuracy_score from keras.models import Sequential from keras.layers import Dense # New Type of Neural Network class GeneticNeuralNetwork(Sequential): # Constructor def __init__(self, child_weights=None): # Initialize Sequential Model Super Class super().__init__() # If no weights provided randomly generate them if child_weights is None: # Layers are created and randomly generated layer1 = Dense(4, input_shape=(4,), activation='sigmoid') layer2 = Dense(2, activation='sigmoid') layer3 = Dense(1, activation='sigmoid') # Layers are added to the model self.add(layer1) self.add(layer2) self.add(layer3) # If weights are provided set them within the layers else: # Set weights within the layers self.add( Dense( 4, input_shape=(4,), activation='sigmoid', weights=[child_weights[0], np.zeros(4)]) ) self.add( Dense( 2, activation='sigmoid', weights=[child_weights[1], np.zeros(2)]) ) self.add( Dense( 1, activation='sigmoid', weights=[child_weights[2], np.zeros(1)]) ) # Function for forward propagating a row vector of a matrix def forward_propagation(self, X_train, y_train): # Forward propagation y_hat = self.predict(X_train.values) # Compute fitness score self.fitness = accuracy_score(y_train, y_hat.round()) # Standard Backpropagation def compile_train(self, epochs): self.compile( optimizer='rmsprop', loss='binary_crossentropy', metrics=['accuracy'] ) self.fit(X_train.values, y_train.values, epochs=epochs)
现在我们已经建立了遗传神经网络,我们可以开发出一种交叉算法。我们将使用类似于上面给出的生物图示的单点交叉。每一个矩阵列都有相同的机会被选择为一个交叉点,让每一个父母组合他们的基因并将它们传递给孩子。
# Crossover traits between two Genetic Neural Networks def dynamic_crossover(nn1, nn2): # Lists for respective weights nn1_weights = [] nn2_weights = [] child_weights = [] # Get all weights from all layers in the first network for layer in nn1.layers: nn1_weights.append(layer.get_weights()[0]) # Get all weights from all layers in the second network for layer in nn2.layers: nn2_weights.append(layer.get_weights()[0]) # Iterate through all weights from all layers for crossover for i in range(0, len(nn1_weights)): # Get single point to split the matrix in parents based on # of cols split = random.randint(0, np.shape(nn1_weights[i])[1]-1) # Iterate through after a single point and set the remaing cols to nn_2 for j in range(split, np.shape(nn1_weights[i])[1]-1): nn1_weights[i][:, j] = nn2_weights[i][:, j] # After crossover add weights to child child_weights.append(nn1_weights[i]) # Add a chance for mutation mutation(child_weights) # Create and return child object child = GeneticNeuralNetwork(child_weights) return child
为了确保种群探索解空间,应该会发生突变。在这种情况下,因为解空间非常大,突变的概率显著高于大多数其他遗传算法。没有特定的方法来改变矩阵,我们在矩阵上随机执行标量乘法,幅度为2-5。
# Chance to mutate weights def mutation(child_weights): # Add a chance for random mutation selection = random.randint(0, len(child_weights)-1) mut = random.uniform(0, 1) if mut >= .5: child_weights[selection] *= random.randint(2, 5) else: # No mutation pass
最后,模拟遗传神经网络的演化。我们需要网络数据来学习,因此我们将使用众所周知的 banknote机器学习数据集。
# Read Data data = pd.read_csv('banknote.csv') # Create Matrix of Independent Variables X = data.drop(['Y'], axis=1) # Create Vector of Dependent Variable y = data['Y'] # Create a Train Test Split for Genetic Optimization X_train, X_test, y_train, y_test = train_test_split(X, y) # Create a List of all active GeneticNeuralNetworks networks = [] pool = [] # Track Generations generation = 0 # Initial Population n = 20 # Generate n randomly weighted neural networks for i in range(0, n): networks.append(GeneticNeuralNetwork()) # Cache Max Fitness max_fitness = 0 # Max Fitness Weights optimal_weights = [] # Evolution Loop while max_fitness < .9: # Log the current generation generation += 1 print('Generation: ', generation) # Forward propagate the neural networks to compute a fitness score for nn in networks: # Propagate to calculate fitness score nn.forward_propagation(X_train, y_train) # Add to pool after calculating fitness pool.append(nn) # Clear for propagation of next children networks.clear() # Sort based on fitness pool = sorted(pool, key=lambda x: x.fitness) pool.reverse() # Find Max Fitness and Log Associated Weights for i in range(0, len(pool)): # If there is a new max fitness among the population if pool[i].fitness > max_fitness: max_fitness = pool[i].fitness print('Max Fitness: ', max_fitness) # Reset optimal_weights optimal_weights = [] # Iterate through all layers, get weights, and append to optimal for layer in pool[i].layers: optimal_weights.append(layer.get_weights()[0]) print(optimal_weights) # Crossover, top 5 randomly select 2 partners for child for i in range(0, 5): for j in range(0, 2): # Create a child and add to networks temp = dynamic_crossover(pool[i], random.choice(pool)) # Add to networks to calculate fitness score next iteration networks.append(temp) # Create a Genetic Neural Network with optimal initial weights gnn = GeneticNeuralNetwork(optimal_weights) gnn.compile_train(10) # Test the Genetic Neural Network Out of Sample y_hat = gnn.predict(X_test.values) print('Test Accuracy: %.2f' % accuracy_score(y_test, y_hat.round()))
第一种模式:10代遗传算法和10个epochs的训练
第二种模式:10个epochs的训练
遗传神经网络在相同数量的训练时期内将模型准确度提高了 0.39。
人工神经网络的初学者很可能都会面临一些问题。人工神经网络的灵感来自于生物学的神经网络。为了简单起见,在计算机科学中,它被表示为一系列的网络层。输入层中的神经元数量等于正在处理的数据中的输入变量的数量。