[PYTHON] Finding the optimum value of a function using a genetic algorithm (Part 1)
What is an evolutionary algorithm?
The evolutionary algorithm (EA) is one of the optimization algorithms, and is an optimization method that incorporates natural selection, gene recombination, mutation, etc. as a model.
Evolutionary algorithms are roughly divided into the following four types. --Genetic algorithm --Genetic programming --Evolution strategy --Evolutionary programming
This time, we will deal with the genetic algorithm (GA).
Library used
Use the following python library called DEAP. https://deap.readthedocs.io/en/master/
code
Objective function
Search for the minimum value of Bird Bunction shown below.
sample_GA.py
import numpy as np
def bird(x):
x1 = x[0]
x2 = x[1]
t = np.sin(x1)*np.exp((1-np.cos(x2))**2) + np.cos(x2)*np.exp((1-np.sin(x1))**2) + (x1-x2)**2
return t,
#drawing
import matplotlib.pyplot as plt
from mpl_toolkits.mplot3d import Axes3D
c1 = c2 = np.arange(-10, 10, 0.1)
xx, yy = np.meshgrid(c1, c2)
len=np.size(c1)
z=np.empty((len, len))
for i in range(len):
for j in range(len):
xin=np.array([xx[i,j], yy[i,j]])
z[i,j], = bird(xin)
fig0 = plt.figure(figsize=(12,12), dpi=80).add_subplot(111, projection='3d')
fig0.plot_surface(xx, yy, z, alpha=1)
plt.xlim(10, -10)
plt.ylim(-10, 10)
Library import
Import the required libraries.
Since DEAP uses the random library, the results change with each trial.
To avoid this, use random.seed (1).
sample_GA.py
import pandas as pd
import random
from deap import algorithms
from deap import base,creator,tools
# random.seed(1)
Early population generation
Create a function to generate an initial individual. Here, initial individuals were generated in a 10 × 10 grid pattern in the range of -10 to 10 for both x and y.
sample_GA.py
def initPopulation(pcls, ind_init):
x = np.linspace(-10, 10, 10)
y = np.linspace(-10, 10, 10)
inix, iniy = np.meshgrid(x,y)
contents = np.concatenate([inix.reshape([-1,1]), iniy.reshape([-1,1])], axis=1)
pcls = [ind_init(c) for c in contents]
return pcls
Register in toolbox
Methods such as evaluation ("evaluate"), population ("population"), crossover ("mate"), mutation ("mutate"), and selection ("select") are registered in the toolbox.
For example, the following ``` toolbox.register (" mate ", tools.cxBlend, alpha = 0.2)` `` It means that the function tool.cxblend is registered as a crossover method (alpha = 0.2 is an argument of tools.cxBlend).
Blend crossing as the crossing method, Gauss mutation as the mutation, and the selection format will be carried out in a tournament of 10 individuals. If you want to specify the initial population, register it in "population_guess".
sample_GA.py
creator.create("FitnessMax", base.Fitness, weights=(-1.0,))
creator.create("Individual", np.ndarray, fitness=creator.FitnessMax)
toolbox = base.Toolbox()
toolbox.register("evaluate", bird)
toolbox.register("population_guess", initPopulation, list, creator.Individual)
toolbox.register("mate", tools.cxBlend, alpha=0.2)
toolbox.register("mutate", tools.mutGaussian, mu=[0.0, 0.0], sigma=[2.0, 2.0], indpb=1)
toolbox.register("select", tools.selTournament, tournsize=10)
Run
If you set up so far, the rest will work with simple commands.
In the example below, the number of evaluated individuals, mean value, standard deviation, maximum value, and minimum value of each generation are recorded in the `` `logbook. hof```Is an abbreviation for holl of fame, and you can save the specified number in descending order of evaluation.
The argument cxpb of algorithms.eaSimple indicates the crossover rate, mutpb indicates the mutation rate, and ngen indicates the number of generations. The number of individuals per generation will be 10 × 10 = 100 individuals, inheriting the one defined by `` `pop```. It is also possible to specify the initial sampling and the number of individuals at the time of generation change separately.
sample_GA.py
#Generate an early population
pop = toolbox.population_guess()
#Setting
hof = tools.HallOfFame(5, similar=np.array_equal)
stats = tools.Statistics(lambda ind: ind.fitness.values)
stats.register("avg", np.mean)
stats.register("std", np.std)
stats.register("min", np.min)
stats.register("max", np.max)
#Run
pop, logbook= algorithms.eaSimple(pop, toolbox,
cxpb=0.9, mutpb=0.1, ngen=20,
stats=stats, halloffame=hof, verbose=True)
#Show top 3 individuals
print("holl of fame.1 : %s" % (hof[0]))
print("holl of fame.2 : %s" % (hof[1]))
print("holl of fame.3 : %s" % (hof[2]))
terminal
gen nevals avg std min max
0 100 80.0665 99.3005 -83.8352 414.98
1 74 7.49092 62.2371 -103.255 301.724
2 89 11.499 108.449 -105.693 665.594
3 84 -60.283 53.2721 -106.71 161.645
4 87 -95.3896 32.299 -106.723 40.9594
5 75 -99.3516 28.4967 -106.758 23.0967
6 73 -104.068 15.7185 -106.764 4.33984
7 76 -103.476 18.6955 -106.764 10.0824
8 80 -101.665 22.5172 -106.764 16.8155
9 92 -102.631 20.6472 -106.764 26.921
10 77 -102.882 19.2791 -106.764 6.34586
11 90 -99.4555 30.4939 -106.764 56.3788
12 89 -100.566 27.1489 -106.764 30.2934
13 79 -100.978 25.2596 -106.764 51.745
14 78 -98.4071 32.1796 -106.764 85.5625
15 76 -105.728 10.3096 -106.764 -3.14868
16 89 -95.2292 38.3427 -106.764 80.6272
17 91 -102.44 25.6436 -106.764 96.6545
18 80 -105.432 11.2501 -106.764 4.33866
19 83 -102.271 23.504 -106.764 67.6912
20 79 -103.856 16.5553 -106.764 -3.86946
holl of fame.1 : [4.70021671 3.1529326 ]
holl of fame.2 : [4.70021671 3.15293287]
holl of fame.3 : [4.70021671 3.15293358]
Globally optimal solution-Converged around 106.765. Since the algorithm does not use the differential information of each individual, a minute error will occur.
Summary and impressions
--Function optimization was performed with the genetic algorithm DEAP. ――Since it uses random numbers, it is doubtful whether it will converge with each trial. -Is it for multivariable optimization with discrete values?
Recommended Posts