### 结合神经网络和遗传算法的Python代码示例 在处理复杂优化问题时，结合神经网络与遗传算法可以提供一种强大的解决方案。下面展示了一个简单的例子，在该例子中使用了遗传算法来优化神经网络权重。 #### 遗传算法简介 遗传算法是一种基于自然选择机制的全局搜索启发式方法，通过模拟生物进化过程中的选择、交叉以及变异操作来进行寻优[^1]。 #### 神经网络结构定义 首先定义一个基本的前馈神经网络模型： ```python import numpy as np class NeuralNetwork(object): def __init__(self, layers=[8, 5, 1], activation='tanh'): self.layers = layers self.num_layers = len(layers) if activation == 'sigmoid': self.activation = lambda x : 1 / (1 + np.exp(-x)) elif activation == 'relu': self.activation = lambda x : np.maximum(x, 0) else: self.activation = lambda x : np.tanh(x) def predict(self, X): for i in range(0, self.num_layers - 1): X = self.activation(np.dot(X, weights[i])) return X ``` 此部分实现了具有指定层数及每层节点数目的多层感知器，并设置了激活函数选项。 #### 定义适应度评估函数 为了使遗传算法能够评价个体的好坏程度，需设计特定于应用场景下的适应度计算方式： ```python def fitness_function(weights, nn_structure, inputs, targets): # 初始化神经网络实例并加载当前染色体对应的权值参数 nn = NeuralNetwork(nn_structure) global_weights_index = 0 for layer_idx in range(len(nn_structure)-1): rows = nn_structure[layer_idx+1] cols = nn_structure[layer_idx] weight_matrix = np.array( weights[ global_weights_index : global_weights_index+(rows*cols) ] ).reshape((rows, cols)) globals()[f'weights_{layer_idx}'] = weight_matrix global_weights_index += rows * cols predictions = nn.predict(inputs).flatten() mse = ((predictions - targets)**2).mean() return mse ``` 这里采用均方误差作为衡量标准，即预测结果越接近目标值，则说明解的质量越高；反之亦然。 #### 实现遗传算法框架 最后构建完整的GA流程，包括种群初始化、迭代更新等环节： ```python from random import Random import matplotlib.pyplot as plt def evolve_population(pop_size=10, mutation_rate=0.01, generations=100, elite_selection_rate=0.2): rng = Random() population = initialize_population(pop_size) best_fitness_history = [] elites_num = int(elite_selection_rate * pop_size) for generation in range(generations): ranked_solutions = rank_by_fitness(population) top_elites = [ solution['chromosome'] for solution in ranked_solutions[:elites_num] ] next_gen = [] while len(next_gen) < pop_size-elites_num: parent_a = select_parent(ranked_solutions, rng) parent_b = select_parent(ranked_solutions, rng) child_chromosome = crossover(parent_a, parent_b, rng) mutate(child_chromosome, mutation_rate, rng) next_gen.append({ "chromosome":child_chromosome, "fitness":-999999 }) new_generation = top_elites + [indv["chromosome"] for indv in next_gen] evaluate_fitness(new_generation) current_best_fit = min([soln['fitness'] for soln in new_generation]) print(f"Generation {generation}: Best Fitness={current_best_fit}") best_fitness_history.append(current_best_fit) population = new_generation.copy() plot_convergence(best_fitness_history) def initialize_population(size): chromosomes = [] total_weights_count = sum([ l_in*l_out for l_in,l_out in zip(nn.layers[:-1],nn.layers[1:]) ]) for _ in range(size): chromosome = { "chromosome":[rng.uniform(-1, 1) for _ in range(total_weights_count)], "fitness":-999999 } chromosomes.append(chromosome) return chromosomes def rank_by_fitness(population): sorted_pop = sorted( population, key=lambda k:k['fitness'], reverse=False ) return sorted_pop def select_parent(sorted_population, rand_obj): index = rand_obj.randint(0,len(sorted_population)//2) return sorted_population[index]['chromosome'] def crossover(a,b,rng): midpoint = rng.randint(0,min(len(a),len(b))-1) offspring = a[:midpoint]+b[midpoint:] return offspring def mutate(individual, rate, rng): for idx,val in enumerate(individual): if(rng.random()<=rate): individual[idx]=rng.uniform(-1,1) def evaluate_fitness(population): for indivdual in population: indivdual['fitness']=fitness_function( indivdual, nn.layers, training_inputs, target_outputs ) def plot_convergence(history): fig=plt.figure(figsize=(7,5)) ax=fig.add_subplot(111) ax.plot(range(len(history)),history,'r-',label="Best Fitness") ax.set_xlabel('Generations') ax.set_ylabel('Fitness Value') ax.legend(loc='upper right', shadow=True) plt.show() if __name__=='__main__': # 假设已知输入输出数据集 training_inputs=np.array([[...]]) target_outputs=np.array([...]) nn = NeuralNetwork([input_dim, hidden_layer_1_units,... ,output_dim]) evolve_population() ``` 上述代码片段展示了如何利用遗传算法训练一个多层感知机模型的过程。需要注意的是这只是一个简化版本用于教学演示用途，在实际应用当中还需要考虑更多细节比如正则化项加入防止过拟合等问题。