差分进化算法
差分进化算法(differential evolution, DE)是一类基于群体的自适应全局优化算法,有较强的鲁棒性和全局寻优能力,主要用于求解实数优化问题。它从数学角度看是一种随机搜索算法,从工程角度看是一种自适应的迭代寻优过程。差分进化算法凭借其独特的优势在数据挖掘、模式识别、电磁学等领域被广泛应用。
算法流程
差分进化算法采用实数编码,基于差分的简单变异操作和“一对一”的竞争生存策略,具体步骤如下
- 随机初始化种群规模N,进化代数t=1;
- 计算初始种群中的每个个体的目标函数值;
- 判断是否达到终止条件或达到最大进化代数。若满足,则进化停止,将此时的最佳个体作为解输出;否则执行步骤4;
- 进行变异操作和交叉操作,并对边界条件进行处理,得到临时种群;
- 对临时种群进行评价,计算临时种群中每个个体的适应度值;
- 进行选择操作,得到新的种群;
- 进化代数t=t+1,并用新的种群代替旧种群,返回步骤3.
伪代码
初始化N个个体
初始化参数D、N_iter
i=1
while(t<=N_iter)
for i=1:N
for j=1:D
进行变异操作和交叉操作,得到临时种群
end for j
计算临时种群中每个个体的适应度值
对临时种群进行选择操作,得到新种群
end for i
t = t+1
end while
输出结果
MATLAB仿真实例
1.计算函数 f ( x ) = ∑ i = 1 n x i 2 ( − 20 ≤ x i ≤ 20 ) f(x)= \sum_{i=1}^n x_i^2(-20\leq x_i\leq20) f(x)=∑i=1nxi2(−20≤xi≤20)的最小值,其中个体 x x x的维数 n = 10 n=10 n=10。这是一个简单的平方和函数,只有一个极小点 x = ( 0 , 0 , ⋯ , 0 ) x=(0,0,\cdots,0) x=(0,0,⋯,0),理论最小值 f ( 0 , 0 , ⋯ , 0 ) = 0 f(0,0,\cdots,0)=0 f(0,0,⋯,0)=0
仿真过程如下:
(1)初始化个体数目为 N P = 50 NP=50 NP=50,变量维数为 D = 10 D=10 D=10,最大进化代数为 G = 200 G=200 G=200,初始变异算子 F 0 = 0.4 F_0=0.4 F0=0.4,交叉算子 C R = 0.1 CR=0.1 CR=0.1,阈值 y z = 1 0 − 6 yz=10^{-6} yz=10−6。
(2)产生初始种群,计算个体目标函数;进行变异操作、交叉操作、边界条件处理,产生临时种群,其中变异操作采用自适应变异算子,边界条件处理采用在可行域中随机产生参数向量的方式。
(3)计算临时种群个体目标函数,与原种群对应个体进行“一对一”选择操作,产生新种群。
(4)判断是否满足终止条件:若满足,则结束搜索过程,输出优化值;若不满足,则继续进行迭代优化。
优化结束后,DE目标函数曲线如下图所示
%% 差分进化算法求极值
clc; clear; close all;
NP = 50; % 种群数量
D = 10; % 变量的维数
G = 200; % 最大的进化代数
F0 = 0.4; % 初始变异算子
CR = 0.1; % 交叉算子
Xs = 20; % 上限
Xx = -20; % 下限
yz = 10^-6; % 阈值
% 赋初值
x = zeros(D, NP); % 初始种群
v = zeros(D, NP); % 变异种群
u = zeros(D, NP); % 选择种群
x = rand(D, NP) * (Xs - Xx) + Xx; % 赋初值
%%%%%%%%%%%%%%计算目标函数%%%%%%%%%%%%%%%%
for m=1:NP
Ob(m) = func1(x(:, m));
end
trace(1) = min(Ob);
%%%%%%%%%%%%%%差分进化循环%%%%%%%%%%%%%%%%
for gen = 1:G
%%%%%%%%%%%%%%%%%%%%%%%%% 变异操作%%%%%%%%%%%%%%%%
%%%%%%%%%%%%%%%%%%%%% 自适应变异算子%%%%%%%%%%%%%%%
lambda = exp(1-G/(G+1-gen));
F = F0*2^(1-lambda);
%%%%%%%%%%%%%%%% r1, r2, r3和m互不相同%%%%%%%%%%%%%%%%
for m = 1:NP
r1 = randi([1, NP], 1, 1);
while(r1==m)
r1 = randi([1, NP], 1, 1);
end
r2 = randi([1, NP], 1, 1);
while(r2==m) || (r2==r1)
r2 = randi([1, NP], 1, 1);
end
r3 = randi([1, NP], 1, 1);
while (r3==m) || (r3==r1) || (r3==r2)
r3 = randi([1, NP], 1, 1);
end
v(:,m) = x(:,r1) + F*(x(:,r2)-x(:,r3));
end
%%%%%%%%%%%%%%%%%%%%%%%交叉操作%%%%%%%%%%%%%%%%%%%%%%%%%%%
r = randi([1, D], 1, 1);
for n = 1:D
cr = rand(1);
if(cr<=CR) || (n==r)
u(n, :) = v(n, :);
else
u(n, :) = x(n, :);
end
end
%%%%%%%%%%%%%%%%%边界条件的处理%%%%%%%%%%%%%%%%%%%%%%%
for n = 1:D
for m = 1:NP
if(u(n,m)<Xx || u(n,m)>Xs)
u(n,m) = rand*(Xs-Xx) + Xx;
end
end
end
%%%%%%%%%%%%%%%%%%%%%选择操作%%%%%%%%%%%%%%%%%%%%%%%%%%
for m = 1:NP
Ob1(m) = func1(u(:, m));
end
for m = 1:NP
if Ob1(m) < Ob(m)
x(:, m) = u(:, m);
end
end
for m = 1:NP
Ob(m) = func1(x(:, m));
end
trace(gen+1) = min(Ob);
if min(Ob(m)) <yz
break
end
end
[SortOb, Index] = sort(Ob);
x = x(:, Index);
X = x(:, 1); % 最优变量
Y = min(Ob); % 最优值
%%%%%%%%%%%%%%%%%%画图%%%%%%%%%%%%%%%%%%
figure
plot(trace);
xlabel('迭代次数');
ylabel('目标函数值');
title('DE目标函数曲线')
%%%%%%%%%%%%适应度函数%%%%%%%%%%%%%%%%%%%
function result = func1(x)
summ = sum(x.^2);
result = summ;
end
- 求函数 f ( x , y ) = 3 cos ( x y ) + x + y f(x,y)=3\cos(xy)+x+y f(x,y)=3cos(xy)+x+y的最小值,其中x的取值范围为[-4,4],y的取值范围为[-4,4]。这是一个有多个局部极值的函数,其函数值图形如下图所示,其MATLAB实现程序如下:
clc; clear; close all;
%%%%%%%%%%%%%%%f(x,y) =3cos(xy)+x+y%%%%%%%%%%%%%%%%
x = -4:0.02:4;
y = -4:0.02:4;
N = size(x, 2);
for i = 1:N
for j = 1:N
z(i,j) = 3*cos(x(i)*y(j))+x(i)+y(j);
end
end
mesh(x, y, z);
xlabel('x');
ylabel('y');
仿真过程如下:
(1)初始化个体数目为NP=20,变量维数为D=2,最大进化代数为G=100,变异算子F=0.5,交叉算子CR=0.1;
(2)产生初始种群,计算个体目标函数;进行变异操作、交叉操作、边界条件处理,产生临时种群,其中边界条件处理采用边界吸收方式;
(3)计算临时种群个体目标函数,与原种群对应个体进行“一对一”选择操作,产生新种群;
(4)判断是否满足终止条件:若满足,则结束搜索过程,输出优化值;若不满足,则继续进行迭代优化。、
优化结束后,DE目标函数曲线如下图所示
clc; clear; close all;
%%%%%%%%%%%%%%%f(x,y) =3cos(xy)+x+y%%%%%%%%%%%%%%%%
x = -4:0.02:4;
y = -4:0.02:4;
N = size(x, 2);
for i = 1:N
for j = 1:N
z(i,j) = 3*cos(x(i)*y(j))+x(i)+y(j);
end
end
mesh(x, y, z);
xlabel('x');
ylabel('y');
%% 差分进化算法求极值
NP = 20; % 种群数量
D = 2; % 变量的维数
G = 100; % 最大的进化代数
F = 0.5; % 初始变异算子
CR = 0.1; % 交叉算子
Xs = 4; % 上限
Xx = -4; % 下限
yz = 10^-6; % 阈值
% 赋初值
x = zeros(D, NP); % 初始种群
v = zeros(D, NP); % 变异种群
u = zeros(D, NP); % 选择种群
x = rand(D, NP) * (Xs - Xx) + Xx; % 赋初值
%%%%%%%%%%%%%%计算目标函数%%%%%%%%%%%%%%%%
for m=1:NP
Ob(m) = func2(x(:, m));
end
trace(1) = min(Ob);
%%%%%%%%%%%%%%差分进化循环%%%%%%%%%%%%%%%%
for gen = 1:G
%%%%%%%%%%%%%%%%%%%%%%%%% 变异操作%%%%%%%%%%%%%%%%
%%%%%%%%%%%%%%%% r1, r2, r3和m互不相同%%%%%%%%%%%%%%%%
for m = 1:NP
r1 = randi([1, NP], 1, 1);
while(r1==m)
r1 = randi([1, NP], 1, 1);
end
r2 = randi([1, NP], 1, 1);
while(r2==m) || (r2==r1)
r2 = randi([1, NP], 1, 1);
end
r3 = randi([1, NP], 1, 1);
while (r3==m) || (r3==r1) || (r3==r2)
r3 = randi([1, NP], 1, 1);
end
v(:,m) = x(:,r1) + F*(x(:,r2)-x(:,r3));
end
%%%%%%%%%%%%%%%%%%%%%%%交叉操作%%%%%%%%%%%%%%%%%%%%%%%%%%%
r = randi([1, D], 1, 1);
for n = 1:D
cr = rand(1);
if(cr<=CR) || (n==r)
u(n, :) = v(n, :);
else
u(n, :) = x(n, :);
end
end
%%%%%%%%%%%%%%%%%边界条件的处理%%%%%%%%%%%%%%%%%%%%%%%
for n = 1:D
for m = 1:NP
if(u(n,m)<Xx)
u(n,m) = Xx;
end
if(u(n,m)>Xs)
u(n,m) = Xs;
end
end
end
%%%%%%%%%%%%%%%%%%%%%选择操作%%%%%%%%%%%%%%%%%%%%%%%%%%
for m = 1:NP
Ob1(m) = func2(u(:, m));
end
for m = 1:NP
if Ob1(m) < Ob(m)
x(:, m) = u(:, m);
end
end
for m = 1:NP
Ob(m) = func2(x(:, m));
end
trace(gen+1) = min(Ob);
% if min(Ob(m)) <yz
% break
% end
end
[SortOb, Index] = sort(Ob);
x = x(:, Index);
X = x(:, 1); % 最优变量
Y = min(Ob); % 最优值
%%%%%%%%%%%%%%%%%%画图%%%%%%%%%%%%%%%%%%
figure
plot(trace);
xlabel('迭代次数');
ylabel('目标函数值');
title('DE目标函数曲线')
%%%%%%%%%%%%适应度函数%%%%%%%%%%%%%%%%%%%
function result = func2(x)
result = 3*cos(x(1)*x(2))+x(1)+x(2);
end
- 用离散差分进化算法求函数 f ( x , y ) = − ( ( x 2 + y − 1 ) 2 + ( x + y 2 − 7 ) 2 ) / 200 + 10 f(x,y)=-((x^2+y-1)^2+(x+y^2-7)^2)/200+10 f(x,y)=−((x2+y−1)2+(x+y2−7)2)/200+10的最大值,其中 x x x的取值为-100~100之间的整数, y y y的取值为-100~100之间的整数,其函数值图形如下图所示
%%%%F(x,y)=-((x~2+y-1).^2+(x+y~2-7)^2)/200+10%%%%
clc; clear; close all;
x=-100:1:100;
y=-100:1:100;
N=size(x,2);
for i=1:N
for j=1:N
z(i,j)=-((x(i)^2+y(j)-1).^2+(x(i) +y(j)^2-7)^2)/200+10;
end
end
mesh(x,y,z);
xlabel('x');
ylabel('y');
仿真过程如下:
(1)初始化个体数目为NP=20,变量维数为D=2,最大进化代数为G=100,变异算子F=0.5,交叉算子CR=0.1
(2)产生数值为[-100,100]内整数的初始种群,计算个体目标函数;进行变异操作,对变异后的种群内数值进行取整操作,然后进行交叉操作、边界条件处理操作,产生临时种群,其中边界条件处理采用边界吸收方式
(3)计算临时种群个体目标函数,与原种群对应个体进行“一对一”选择操作,产生新种群
(4)判断是否满足终止条件:若满足,则结束搜索过程,输出优化值;若不满足,则继续进行迭代优化
优化结束后,DE目标函数曲线如下图所示
%%%%F(x,y)=-((x~2+y-1).^2+(x+y~2-7)^2)/200+10%%%%
clc; clear; close all;
x=-100:1:100;
y=-100:1:100;
N=size(x,2);
for i=1:N
for j=1:N
z(i,j)=-((x(i)^2+y(j)-1).^2+(x(i) +y(j)^2-7)^2)/200+10;
end
end
mesh(x,y,z);
xlabel('x');
ylabel('y');
%% 差分进化算法求极值
NP = 20; % 种群数量
D = 2; % 变量的维数
G = 100; % 最大的进化代数
F = 0.5; % 初始变异算子
CR = 0.1; % 交叉算子
Xs = 100; % 上限
Xx = -100; % 下限
yz = 10^-6; % 阈值
% 赋初值
x = zeros(D, NP); % 初始种群
v = zeros(D, NP); % 变异种群
u = zeros(D, NP); % 选择种群
x = randi([Xx, Xs], D, NP); % 赋初值
%%%%%%%%%%%%%%计算目标函数%%%%%%%%%%%%%%%%
for m=1:NP
Ob(m) = func3(x(:, m));
end
trace(1) = max(Ob);
%%%%%%%%%%%%%%差分进化循环%%%%%%%%%%%%%%%%
for gen = 1:G
%%%%%%%%%%%%%%%%%%%%%%%%% 变异操作%%%%%%%%%%%%%%%%
%%%%%%%%%%%%%%%% r1, r2, r3和m互不相同%%%%%%%%%%%%%%%%
for m = 1:NP
r1 = randi([1, NP], 1, 1);
while(r1==m)
r1 = randi([1, NP], 1, 1);
end
r2 = randi([1, NP], 1, 1);
while(r2==m) || (r2==r1)
r2 = randi([1, NP], 1, 1);
end
r3 = randi([1, NP], 1, 1);
while (r3==m) || (r3==r1) || (r3==r2)
r3 = randi([1, NP], 1, 1);
end
v(:,m) = floor(x(:,r1) + F*(x(:,r2)-x(:,r3)));
end
%%%%%%%%%%%%%%%%%%%%%%%交叉操作%%%%%%%%%%%%%%%%%%%%%%%%%%%
r = randi([1, D], 1, 1);
for n = 1:D
cr = rand(1);
if(cr<=CR) || (n==r)
u(n, :) = v(n, :);
else
u(n, :) = x(n, :);
end
end
%%%%%%%%%%%%%%%%%边界条件的处理%%%%%%%%%%%%%%%%%%%%%%%
for n = 1:D
for m = 1:NP
if(u(n,m)<Xx)
u(n,m) = Xx;
end
if(u(n,m)>Xs)
u(n,m) = Xs;
end
end
end
%%%%%%%%%%%%%%%%%%%%%选择操作%%%%%%%%%%%%%%%%%%%%%%%%%%
for m = 1:NP
Ob1(m) = func3(u(:, m));
end
for m = 1:NP
if Ob1(m) > Ob(m)
x(:, m) = u(:, m);
end
end
for m = 1:NP
Ob(m) = func3(x(:, m));
end
trace(gen+1) = max(Ob);
% if min(Ob(m)) <yz
% break
% end
end
[SortOb, Index] = sort(Ob);
x = x(:, Index);
X = x(:, end); % 最优变量
Y = max(Ob); % 最优值
%%%%%%%%%%%%%%%%%%画图%%%%%%%%%%%%%%%%%%
figure
plot(trace);
xlabel('迭代次数');
ylabel('目标函数值');
title('DE目标函数曲线')
%%%%%%%%%%%%适应度函数%%%%%%%%%%%%%%%%%%%
function result = func3(x)
result = -((x(1)^2+x(2)-1).^2+(x(1) +x(2)^2-7)^2)/200+10;
end
参考
[1] 包子阳等. 智能优化算法及其MATLAB实例(第3版)[M]. 北京: 电子工业出版社, 2020.