您好,登錄后才能下訂單哦!
小編給大家分享一下python如何實現(xiàn)高效的遺傳算法,相信大部分人都還不怎么了解,因此分享這篇文章給大家參考一下,希望大家閱讀完這篇文章后大有收獲,下面讓我們一起去了解一下吧!
遺傳算法屬于一種優(yōu)化算法。
如果你有一個待優(yōu)化函數(shù),可以考慮次算法。假設你有一個變量x,通過某個函數(shù)可以求出對應的y,那么你通過預設的x可求出y_pred,y_pred差距與你需要的y當然越接近越好,這就需要引入適應度(fitness)的概念。假設
fitness = 1/(1+ads(y_pred - y)),那么誤差越小,適應度越大,即該個體越易于存活。
設計該算法的思路如下:
(1)初始化種群,即在我需要的區(qū)間如[-100,100]內(nèi)random一堆初始個體[x1,x2,x3...],這些個體是10進制形式的,為了后面的交叉與變異我們不妨將其轉(zhuǎn)化為二進制形式。那么現(xiàn)在的問題是二進制取多少位合適呢?即編碼(code)的長度是多少呢?
這就涉及一些信號方面的知識,比如兩位的二進制表示的最大值是3(11),可以將區(qū)間化為4分,那么每一份區(qū)間range長度range/4,我們只需要讓range/n小于我們定義的精度即可。n是二進制需要表示的最大,可以反解出二進制位數(shù) 。
(2)我們需要編寫編碼與解碼函數(shù)。即code:將x1,x2...化為二進制,decode:在交叉變異后重新得到十進制數(shù),用于計算fitness。
(3)交叉后變異函數(shù)編寫都很簡單,random一個point,指定兩個x在point位置進行切片交換即是交叉。變異也是random一個point,讓其值0變?yōu)?,1變?yōu)?。
(4)得到交叉變異后的個體,需要計算fitness進行種群淘汰,保留fitness最高的一部分種群。
(5)將最優(yōu)的個體繼續(xù)上面的操作,直到你定義的iteration結(jié)束為止。
不說了,上代碼:
import numpy as np import pandas as pd import random from scipy.optimize import fsolve import matplotlib.pyplot as plt import heapq from sklearn.model_selection import train_test_split from tkinter import _flatten from sklearn.utils import shuffle from sklearn import preprocessing from sklearn.decomposition import PCA from matplotlib import rcParams # 求染色體長度 def getEncodeLength(decisionvariables, delta): # 將每個變量的編碼長度放入數(shù)組 lengths = [] for decisionvar in decisionvariables: uper = decisionvar[1] low = decisionvar[0] # res()返回一個數(shù)組 res = fsolve(lambda x: ((uper - low) / delta - 2 ** x + 1), 30) # ceil()向上取整 length = int(np.ceil(res[0])) lengths.append(length) # print("染色體長度:", lengths) return lengths # 隨機生成初始化種群 def getinitialPopulation(length, populationSize): chromsomes = np.zeros((populationSize, length), dtype=np.int) for popusize in range(populationSize): # np.random.randit()產(chǎn)生[0,2)之間的隨機整數(shù),第三個參數(shù)表示隨機數(shù)的數(shù)量 chromsomes[popusize, :] = np.random.randint(0, 2, length) return chromsomes # 染色體解碼得到表現(xiàn)形的解 def getDecode(population, encodelength, decisionvariables, delta): # 得到population中有幾個元素 populationsize = population.shape[0] length = len(encodelength) decodeVariables = np.zeros((populationsize, length), dtype=np.float) # 將染色體拆分添加到解碼數(shù)組decodeVariables中 for i, populationchild in enumerate(population): # 設置起始點 start = 0 for j, lengthchild in enumerate(encodelength): power = lengthchild - 1 decimal = 0 start_end = start + lengthchild for k in range(start, start_end): # 二進制轉(zhuǎn)為十進制 decimal += populationchild[k] * (2 ** power) power = power - 1 # 從下一個染色體開始 start = start_end lower = decisionvariables[j][0] uper = decisionvariables[j][1] # 轉(zhuǎn)換為表現(xiàn)形 decodevalue = lower + decimal * (uper - lower) / (2 ** lengthchild - 1) # 將解添加到數(shù)組中 decodeVariables[i][j] = decodevalue return decodeVariables # 選擇新的種群 def selectNewPopulation(decodepopu, cum_probability): # 獲取種群的規(guī)模和 m, n = decodepopu.shape # 初始化新種群 newPopulation = np.zeros((m, n)) for i in range(m): # 產(chǎn)生一個0到1之間的隨機數(shù) randomnum = np.random.random() # 輪盤賭選擇 for j in range(m): if (randomnum < cum_probability[j]): newPopulation[i] = decodepopu[j] break return newPopulation # 新種群交叉 def crossNewPopulation(newpopu, prob): m, n = newpopu.shape # uint8將數(shù)值轉(zhuǎn)換為無符號整型 numbers = np.uint8(m * prob) # 如果選擇的交叉數(shù)量為奇數(shù),則數(shù)量加1 if numbers % 2 != 0: numbers = numbers + 1 # 初始化新的交叉種群 updatepopulation = np.zeros((m, n), dtype=np.uint8) # 隨機生成需要交叉的染色體的索引號 index = random.sample(range(m), numbers) # 不需要交叉的染色體直接復制到新的種群中 for i in range(m): if not index.__contains__(i): updatepopulation[i] = newpopu[i] # 交叉操作 j = 0 while j < numbers: # 隨機生成一個交叉點,np.random.randint()返回的是一個列表 crosspoint = np.random.randint(0, n, 1) crossPoint = crosspoint[0] # a = index[j] # b = index[j+1] updatepopulation[index[j]][0:crossPoint] = newpopu[index[j]][0:crossPoint] updatepopulation[index[j]][crossPoint:] = newpopu[index[j + 1]][crossPoint:] updatepopulation[index[j + 1]][0:crossPoint] = newpopu[j + 1][0:crossPoint] updatepopulation[index[j + 1]][crossPoint:] = newpopu[index[j]][crossPoint:] j = j + 2 return updatepopulation # 變異操作 def mutation(crosspopulation, mutaprob): # 初始化變異種群 mutationpopu = np.copy(crosspopulation) m, n = crosspopulation.shape # 計算需要變異的基因數(shù)量 mutationnums = np.uint8(m * n * mutaprob) # 隨機生成變異基因的位置 mutationindex = random.sample(range(m * n), mutationnums) # 變異操作 for geneindex in mutationindex: # np.floor()向下取整返回的是float型 row = np.uint8(np.floor(geneindex / n)) colume = geneindex % n if mutationpopu[row][colume] == 0: mutationpopu[row][colume] = 1 else: mutationpopu[row][colume] = 0 return mutationpopu # 找到重新生成的種群中適應度值最大的染色體生成新種群 def findMaxPopulation(population, maxevaluation, maxSize): #將數(shù)組轉(zhuǎn)換為列表 #maxevalue = maxevaluation.flatten() maxevaluelist = maxevaluation # 找到前100個適應度最大的染色體的索引 maxIndex = map(maxevaluelist.index, heapq.nlargest(maxSize, maxevaluelist)) index = list(maxIndex) colume = population.shape[1] # 根據(jù)索引生成新的種群 maxPopulation = np.zeros((maxSize, colume)) i = 0 for ind in index: maxPopulation[i] = population[ind] i = i + 1 return maxPopulation # 得到每個個體的適應度值及累計概率 def getFitnessValue(decode,x_train,y_train): # 得到種群的規(guī)模和決策變量的個數(shù) popusize, decisionvar = decode.shape fitnessValue = [] for j in range(len(decode)): W1 = decode[j][0:20].reshape(4,5) V1 = decode[j][20:25].T W2 = decode[j][25:45].reshape(5,4) V2 = decode[j][45:].T error_all = [] for i in range(len(x_train)): #get values of hidde layer X2 = sigmoid(x_train[i].T.dot(W1)+V1) #get values of prediction y Y_hat = sigmoid(X2.T.dot(W2)+V2) #get error when input dimension is i error = sum(abs(Y_hat - y_train[i])) error_all.append(error) #get fitness when W and V is j fitnessValue.append(1/(1+sum(error_all))) # 得到每個個體被選擇的概率 probability = fitnessValue / np.sum(fitnessValue) # 得到每個染色體被選中的累積概率,用于輪盤賭算子使用 cum_probability = np.cumsum(probability) return fitnessValue, cum_probability def getFitnessValue_accuracy(decode,x_train,y_train): # 得到種群的規(guī)模和決策變量的個數(shù) popusize, decisionvar = decode.shape fitnessValue = [] for j in range(len(decode)): W1 = decode[j][0:20].reshape(4,5) V1 = decode[j][20:25].T W2 = decode[j][25:45].reshape(5,4) V2 = decode[j][45:].T accuracy = [] for i in range(len(x_train)): #get values of hidde layer X2 = sigmoid(x_train[i].T.dot(W1)+V1) #get values of prediction y Y_hat = sigmoid(X2.T.dot(W2)+V2) #get error when input dimension is i accuracy.append(sum(abs(np.round(Y_hat) - y_train[i]))) fitnessValue.append(sum([m == 0 for m in accuracy])/len(accuracy)) # 得到每個個體被選擇的概率 probability = fitnessValue / np.sum(fitnessValue) # 得到每個染色體被選中的累積概率,用于輪盤賭算子使用 cum_probability = np.cumsum(probability) return fitnessValue, cum_probability def getXY(): # 要打開的文件名 data_set = pd.read_csv('all-bp.csv', header=None) # 取出“特征”和“標簽”,并做了轉(zhuǎn)置,將列轉(zhuǎn)置為行 X_minMax1 = data_set.iloc[:, 0:12].values # 前12列是特征 min_max_scaler = preprocessing.MinMaxScaler() X_minMax = min_max_scaler.fit_transform(X_minMax1) # 0-1 range transfer = PCA(n_components=0.9) data1 = transfer.fit_transform(X_minMax) #print('PCA processed shape:',data1.shape) X = data1 Y = data_set.iloc[ : , 12:16].values # 后3列是標簽 # 分訓練和測試集 x_train, x_test, y_train, y_test = train_test_split(X, Y, test_size=0.3) return x_train, x_test, y_train, y_test def sigmoid(z): return 1 / (1 + np.exp(-z))
上面的計算適應度函數(shù)需要自己更具實際情況調(diào)整。
optimalvalue = [] optimalvariables = [] # 兩個決策變量的上下界,多維數(shù)組之間必須加逗號 decisionVariables = [[-100,100]]*49 # 精度 delta = 0.001 # 獲取染色體長度 EncodeLength = getEncodeLength(decisionVariables, delta) # 種群數(shù)量 initialPopuSize = 100 # 初始生成100個種群,20,5,20,4分別對用W1,V1,W2,V2 population = getinitialPopulation(sum(EncodeLength), initialPopuSize) print("polpupation.shape:",population.shape) # 最大進化代數(shù) maxgeneration = 4000 # 交叉概率 prob = 0.8 # 變異概率 mutationprob = 0.5 # 新生成的種群數(shù)量 maxPopuSize = 30 x_train, x_test, y_train, y_test = getXY() for generation in range(maxgeneration): # 對種群解碼得到表現(xiàn)形 print(generation) decode = getDecode(population, EncodeLength, decisionVariables, delta) #print('the shape of decode:',decode.shape # 得到適應度值和累計概率值 evaluation, cum_proba = getFitnessValue_accuracy(decode,x_train,y_train) # 選擇新的種群 newpopulations = selectNewPopulation(population, cum_proba) # 新種群交叉 crossPopulations = crossNewPopulation(newpopulations, prob) # 變異操作 mutationpopulation = mutation(crossPopulations, mutationprob) # 將父母和子女合并為新的種群 totalpopulation = np.vstack((population, mutationpopulation)) # 最終解碼 final_decode = getDecode(totalpopulation, EncodeLength, decisionVariables, delta) # 適應度評估 final_evaluation, final_cumprob = getFitnessValue_accuracy(final_decode,x_train,y_train) #選出適應度最大的100個重新生成種群 population = findMaxPopulation(totalpopulation, final_evaluation, maxPopuSize) # 找到本輪中適應度最大的值 optimalvalue.append(np.max(final_evaluation)) index = np.where(final_evaluation == max(final_evaluation)) optimalvariables.append(list(final_decode[index[0][0]]))
fig = plt.figure(dpi = 160,figsize=(5,4)) config = { "font.family":"serif", #serif "font.size": 10, "mathtext.fontset":'stix', } rcParams.update(config) plt.plot(np.arange(len(optimalvalue)), optimalvalue, color="y", lw=0.8, ls='-', marker='o', ms=8) # 圖例設置 plt.xlabel('Iteration') plt.ylabel('Accuracy') plt.show()
以上是“python如何實現(xiàn)高效的遺傳算法”這篇文章的所有內(nèi)容,感謝各位的閱讀!相信大家都有了一定的了解,希望分享的內(nèi)容對大家有所幫助,如果還想學習更多知識,歡迎關注億速云行業(yè)資訊頻道!
免責聲明:本站發(fā)布的內(nèi)容(圖片、視頻和文字)以原創(chuàng)、轉(zhuǎn)載和分享為主,文章觀點不代表本網(wǎng)站立場,如果涉及侵權請聯(lián)系站長郵箱:is@yisu.com進行舉報,并提供相關證據(jù),一經(jīng)查實,將立刻刪除涉嫌侵權內(nèi)容。