该算法简单易懂,很适合刚入门多目标算法的。且想改进其他单目标优化算法为多目标的,完全可以在此算法框架上直接修改!今天重新推出一下。
多目标粒子群算法是应用最广泛,也是最经典的多目标寻优算法。各种硕士博士文章,都将其应用在各种各样的领域。今天就为大家带来一期多目标粒子群算法。
与网上大多数多目标粒子群代码不同,本期给出的多目标粒子群优化算法,只有一个脚本和一个函数,很适合新手学习,而且出图精美!
在经典的多目标测试函数“ZDT1”,“ZDT2”,“ZDT3”,“ZDT6”,“Kursawe”,“Schaffer”,“Poloni”,“Viennet2”,“Viennet3”中对多目标粒子群进行测试,结果如下:
其中绿色的线代表真实的Pareto前沿面,黑色的圆圈表示寻优得到的Pareto值,红色的圈表示其他粒子。
ZDT1
ZDT2
ZDT3
ZDT6
Kursawe
Schaffer
Viennet2
Viennet3
可以看到,在这几个经典函数中的测试,多目标粒子群的效果还是非常不错的,但也有可改进的空间。
接下来直接上代码!
clear all; clc;
% Multi-objective function
% MultiObjFnc = 'Schaffer';
% MultiObjFnc = 'Kursawe';
% MultiObjFnc = 'Poloni';
% MultiObjFnc = 'Viennet2';
% MultiObjFnc = 'Viennet3';
% MultiObjFnc = 'ZDT1';
% MultiObjFnc = 'ZDT2';
% MultiObjFnc = 'ZDT3';
% MultiObjFnc = 'ZDT6';
switch MultiObjFnc
case 'Schaffer' % Schaffer
MultiObj.fun = @(x) [x(:).^2, (x(:)-2).^2];
MultiObj.nVar = 1;
MultiObj.var_min = -5;
MultiObj.var_max = 5;
load('Schaffer.mat');
MultiObj.truePF = PF;
case 'Kursawe' % Kursawe
MultiObj.fun = @(x) [-10.*(exp(-0.2.*sqrt(x(:,1).^2+x(:,2).^2)) + exp(-0.2.*sqrt(x(:,2).^2+x(:,3).^2))), ...
sum(abs(x).^0.8 + 5.*sin(x.^3),2)];
MultiObj.nVar = 3;
MultiObj.var_min = -5.*ones(1,MultiObj.nVar);
MultiObj.var_max = 5.*ones(1,MultiObj.nVar);
load('Kursawe.mat');
MultiObj.truePF = PF;
case 'Poloni' % Poloni's two-objective
A1 = 0.5*sin(1)-2*cos(1)+sin(2)-1.5*cos(2);
A2 = 1.5*sin(1)-cos(1)+2*sin(2)-0.5*cos(2);
B1 = @(x,y) 0.5.*sin(x)-2.*cos(x)+sin(y)-1.5.*cos(y);
B2 = @(x,y) 1.5.*sin(x)-cos(x)+2.*sin(y)-0.5.*cos(y);
f1 = @(x,y) 1+(A1-B1(x,y)).^2+(A2-B2(x,y)).^2;
f2 = @(x,y) (x+3).^2+(y+1).^2;
MultiObj.fun = @(x) [f1(x(:,1),x(:,2)), f2(x(:,1),x(:,2))];
MultiObj.nVar = 2;
MultiObj.var_min = -pi.*ones(1,MultiObj.nVar);
MultiObj.var_max = pi.*ones(1,MultiObj.nVar);
case 'Viennet2' % Viennet2
f1 = @(x,y) 0.5.*(x-2).^2+(1/13).*(y+1).^2+3;
f2 = @(x,y) (1/36).*(x+y-3).^2+(1/8).*(-x+y+2).^2-17;
f3 = @(x,y) (1/175).*(x+2.*y-1).^2+(1/17).*(2.*y-x).^2-13;
MultiObj.fun = @(x) [f1(x(:,1),x(:,2)), f2(x(:,1),x(:,2)), f3(x(:,1),x(:,2))];
MultiObj.nVar = 2;
MultiObj.var_min = [-4, -4];
MultiObj.var_max = [4, 4];
load('Viennet2.mat');
MultiObj.truePF = PF;
case 'Viennet3' % Viennet3
f1 = @(x,y) 0.5.*(x.^2+y.^2)+sin(x.^2+y.^2);
f2 = @(x,y) (1/8).*(3.*x-2.*y+4).^2 + (1/27).*(x-y+1).^2 +15;
f3 = @(x,y) (1./(x.^2+y.^2+1))-1.1.*exp(-(x.^2+y.^2));
MultiObj.fun = @(x) [f1(x(:,1),x(:,2)), f2(x(:,1),x(:,2)), f3(x(:,1),x(:,2))];
MultiObj.nVar = 2;
MultiObj.var_min = [-3, -10];
MultiObj.var_max = [10, 3];
load('Viennet3.mat');
MultiObj.truePF = PF;
case 'ZDT1' % ZDT1 (convex)
g = @(x) 1+9.*sum(x(:,2:end),2)./(size(x,2)-1);
MultiObj.fun = @(x) [x(:,1), g(x).*(1-sqrt(x(:,1)./g(x)))];
MultiObj.nVar = 30;
MultiObj.var_min = zeros(1,MultiObj.nVar);
MultiObj.var_max = ones(1,MultiObj.nVar);
load('ZDT1.mat');
MultiObj.truePF = PF;
case 'ZDT2' % ZDT2 (non-convex)
f = @(x) x(:,1);
g = @(x) 1+9.*sum(x(:,2:end),2)./(size(x,2)-1);
h = @(x) 1-(f(x)./g(x)).^2;
MultiObj.fun = @(x) [f(x), g(x).*h(x)];
MultiObj.nVar = 30;
MultiObj.var_min = zeros(1,MultiObj.nVar);
MultiObj.var_max = ones(1,MultiObj.nVar);
load('ZDT2.mat');
MultiObj.truePF = PF;
case 'ZDT3' % ZDT3 (discrete)
f = @(x) x(:,1);
g = @(x) 1+(9/size(x,2)-1).*sum(x(:,2:end),2);
h = @(x) 1 - sqrt(f(x)./g(x)) - (f(x)./g(x)).*sin(10.*pi.*f(x));
MultiObj.fun = @(x) [f(x), g(x).*h(x)];
MultiObj.nVar = 5;
MultiObj.var_min = 0.*ones(1,MultiObj.nVar);
MultiObj.var_max = 1.*ones(1,MultiObj.nVar);
load('ZDT3.mat');
MultiObj.truePF = PF;
case 'ZDT6' % ZDT6 (non-uniform)
f = @(x) 1 - exp(-4.*x(:,1)).*sin(6.*pi.*x(:,1));
g = @(x) 1 + 9.*(sum(x(:,2:end),2)./(size(x,2)-1)).^0.25;
h = @(x) 1 - (f(x)./g(x)).^2;
MultiObj.fun = @(x) [f(x), g(x).*h(x)];
MultiObj.nVar = 10;
MultiObj.var_min = 0.*ones(1,MultiObj.nVar);
MultiObj.var_max = 1.*ones(1,MultiObj.nVar);
load('ZDT6.mat');
MultiObj.truePF = PF;
end
% Parameters
params.Np = 200; % Population size
params.Nr = 200; % Repository size
params.maxgen = 100; % Maximum number of generations
params.W = 0.4; % Inertia weight
params.C1 = 2; % Individual confidence factor
params.C2 = 2; % Swarm confidence factor
params.ngrid = 20; % Number of grids in each dimension
params.maxvel = 5; % Maxmium vel in percentage
params.u_mut = 0.5; % Uniform mutation percentage
% MOPSO
REP = MOPSO(params,MultiObj);
% Display info
display('Repository fitness values are stored in REP.pos_fit');
display('Repository particles positions are store in REP.pos');
function REP = MOPSO(params,MultiObj)
% Parameters
Np = params.Np;
Nr = params.Nr;
maxgen = params.maxgen;
W = params.W;
C1 = params.C1;
C2 = params.C2;
ngrid = params.ngrid;
maxvel = params.maxvel;
u_mut = params.u_mut;
fun = MultiObj.fun;
nVar = MultiObj.nVar;
var_min = MultiObj.var_min(:);
var_max = MultiObj.var_max(:);
% Initialization
POS = repmat((var_max-var_min)',Np,1).*rand(Np,nVar) + repmat(var_min',Np,1);
VEL = zeros(Np,nVar);
POS_fit = fun(POS);
if size(POS,1) ~= size(POS_fit,1)
warning(['The objective function is badly programmed. It is not returning' ...
'a value for each particle, please check it.']);
end
PBEST = POS;
PBEST_fit= POS_fit;
DOMINATED= checkDomination(POS_fit);
REP.pos = POS(~DOMINATED,:);
REP.pos_fit = POS_fit(~DOMINATED,:);
REP = updateGrid(REP,ngrid);
maxvel = (var_max-var_min).*maxvel./100;
gen = 1;
% Plotting and verbose
if(size(POS_fit,2)==2)
h_fig = figure(1);
h_par = plot(POS_fit(:,1),POS_fit(:,2),'or'); hold on;
h_rep = plot(REP.pos_fit(:,1),REP.pos_fit(:,2),'ok'); hold on;
try
set(gca,'xtick',REP.hypercube_limits(:,1)','ytick',REP.hypercube_limits(:,2)');
axis([min(REP.hypercube_limits(:,1)) max(REP.hypercube_limits(:,1)) ...
min(REP.hypercube_limits(:,2)) max(REP.hypercube_limits(:,2))]);
grid on; xlabel('f1'); ylabel('f2');
end
drawnow;
end
if(size(POS_fit,2)==3)
h_fig = figure(1);
h_par = plot3(POS_fit(:,1),POS_fit(:,2),POS_fit(:,3),'or'); hold on;
h_rep = plot3(REP.pos_fit(:,1),REP.pos_fit(:,2),REP.pos_fit(:,3),'ok'); hold on;
try
set(gca,'xtick',REP.hypercube_limits(:,1)','ytick',REP.hypercube_limits(:,2)','ztick',REP.hypercube_limits(:,3)');
axis([min(REP.hypercube_limits(:,1)) max(REP.hypercube_limits(:,1)) ...
min(REP.hypercube_limits(:,2)) max(REP.hypercube_limits(:,2))]);
end
grid on; xlabel('f1'); ylabel('f2'); zlabel('f3');
drawnow;
axis square;
end
display(['Generation #0 - Repository size: ' num2str(size(REP.pos,1))]);
% Main MPSO loop
stopCondition = false;
while ~stopCondition
% Select leader
h = selectLeader(REP);
% Update speeds and positions
VEL = W.*VEL + C1*rand(Np,nVar).*(PBEST-POS) ...
+ C2*rand(Np,nVar).*(repmat(REP.pos(h,:),Np,1)-POS);
POS = POS + VEL;
% Perform mutation
POS = mutation(POS,gen,maxgen,Np,var_max,var_min,nVar,u_mut);
% Check boundaries
[POS,VEL] = checkBoundaries(POS,VEL,maxvel,var_max,var_min);
% Evaluate the population
POS_fit = fun(POS);
% Update the repository
REP = updateRepository(REP,POS,POS_fit,ngrid);
if(size(REP.pos,1)>Nr)
REP = deleteFromRepository(REP,size(REP.pos,1)-Nr,ngrid);
end
% Update the best positions found so far for each particle
pos_best = dominates(POS_fit, PBEST_fit);
best_pos = ~dominates(PBEST_fit, POS_fit);
best_pos(rand(Np,1)>=0.5) = 0;
if(sum(pos_best)>1)
PBEST_fit(pos_best,:) = POS_fit(pos_best,:);
PBEST(pos_best,:) = POS(pos_best,:);
end
if(sum(best_pos)>1)
PBEST_fit(best_pos,:) = POS_fit(best_pos,:);
PBEST(best_pos,:) = POS(best_pos,:);
end
% Plotting and verbose
if(size(POS_fit,2)==2)
figure(h_fig); delete(h_par); delete(h_rep);
h_par = plot(POS_fit(:,1),POS_fit(:,2),'or'); hold on;
h_rep = plot(REP.pos_fit(:,1),REP.pos_fit(:,2),'ok'); hold on;
try
set(gca,'xtick',REP.hypercube_limits(:,1)','ytick',REP.hypercube_limits(:,2)');
axis([min(REP.hypercube_limits(:,1)) max(REP.hypercube_limits(:,1)) ...
min(REP.hypercube_limits(:,2)) max(REP.hypercube_limits(:,2))]);
end
if(isfield(MultiObj,'truePF'))
try delete(h_pf); end
h_pf = plot(MultiObj.truePF(:,1),MultiObj.truePF(:,2),'.','color','g'); hold on;
end
grid on; xlabel('f1'); ylabel('f2');
drawnow;
axis square;
end
if(size(POS_fit,2)==3)
figure(h_fig); delete(h_par); delete(h_rep);
h_par = plot3(POS_fit(:,1),POS_fit(:,2),POS_fit(:,3),'or'); hold on;
h_rep = plot3(REP.pos_fit(:,1),REP.pos_fit(:,2),REP.pos_fit(:,3),'ok'); hold on;
try
set(gca,'xtick',REP.hypercube_limits(:,1)','ytick',REP.hypercube_limits(:,2)','ztick',REP.hypercube_limits(:,3)');
axis([min(REP.hypercube_limits(:,1)) max(REP.hypercube_limits(:,1)) ...
min(REP.hypercube_limits(:,2)) max(REP.hypercube_limits(:,2)) ...
min(REP.hypercube_limits(:,3)) max(REP.hypercube_limits(:,3))]);
end
if(isfield(MultiObj,'truePF'))
try delete(h_pf); end
h_pf = plot3(MultiObj.truePF(:,1),MultiObj.truePF(:,2),MultiObj.truePF(:,3),'.','color','g'); hold on;
end
grid on; xlabel('f1'); ylabel('f2'); zlabel('f3');
drawnow;
axis square;
end
display(['Generation #' num2str(gen) ' - Repository size: ' num2str(size(REP.pos,1))]);
% Update generation and check for termination
gen = gen + 1;
if(gen>maxgen), stopCondition = true; end
end
hold off;
end
% Function that updates the repository given a new population and its
% fitness
function REP = updateRepository(REP,POS,POS_fit,ngrid)
% Domination between particles
DOMINATED = checkDomination(POS_fit);
REP.pos = [REP.pos; POS(~DOMINATED,:)];
REP.pos_fit= [REP.pos_fit; POS_fit(~DOMINATED,:)];
% Domination between nondominated particles and the last repository
DOMINATED = checkDomination(REP.pos_fit);
REP.pos_fit= REP.pos_fit(~DOMINATED,:);
REP.pos = REP.pos(~DOMINATED,:);
% Updating the grid
REP = updateGrid(REP,ngrid);
end
% Function that corrects the positions and velocities of the particles that
% exceed the boundaries
function [POS,VEL] = checkBoundaries(POS,VEL,maxvel,var_max,var_min)
% Useful matrices
Np = size(POS,1);
MAXLIM = repmat(var_max(:)',Np,1);
MINLIM = repmat(var_min(:)',Np,1);
MAXVEL = repmat(maxvel(:)',Np,1);
MINVEL = repmat(-maxvel(:)',Np,1);
% Correct positions and velocities
VEL(VEL>MAXVEL) = MAXVEL(VEL>MAXVEL);
VEL(VEL<MINVEL) = MINVEL(VEL<MINVEL);
VEL(POS>MAXLIM) = (-1).*VEL(POS>MAXLIM);
POS(POS>MAXLIM) = MAXLIM(POS>MAXLIM);
VEL(POS<MINLIM) = (-1).*VEL(POS<MINLIM);
POS(POS<MINLIM) = MINLIM(POS<MINLIM);
end
% Function for checking the domination between the population. It
% returns a vector that indicates if each particle is dominated (1) or not
function dom_vector = checkDomination(fitness)
Np = size(fitness,1);
dom_vector = zeros(Np,1);
all_perm = nchoosek(1:Np,2); % Possible permutations
all_perm = [all_perm; [all_perm(:,2) all_perm(:,1)]];
d = dominates(fitness(all_perm(:,1),:),fitness(all_perm(:,2),:));
dominated_particles = unique(all_perm(d==1,2));
dom_vector(dominated_particles) = 1;
end
% Function that returns 1 if x dominates y and 0 otherwise
function d = dominates(x,y)
d = all(x<=y,2) & any(x<y,2);
end
% Function that updates the hypercube grid, the hypercube where belongs
% each particle and its quality based on the number of particles inside it
function REP = updateGrid(REP,ngrid)
% Computing the limits of each hypercube
ndim = size(REP.pos_fit,2);
REP.hypercube_limits = zeros(ngrid+1,ndim);
for dim = 1:1:ndim
REP.hypercube_limits(:,dim) = linspace(min(REP.pos_fit(:,dim)),max(REP.pos_fit(:,dim)),ngrid+1)';
end
% Computing where belongs each particle
npar = size(REP.pos_fit,1);
REP.grid_idx = zeros(npar,1);
REP.grid_subidx = zeros(npar,ndim);
for n = 1:1:npar
idnames = [];
for d = 1:1:ndim
REP.grid_subidx(n,d) = find(REP.pos_fit(n,d)<=REP.hypercube_limits(:,d)',1,'first')-1;
if(REP.grid_subidx(n,d)==0), REP.grid_subidx(n,d) = 1; end
idnames = [idnames ',' num2str(REP.grid_subidx(n,d))];
end
REP.grid_idx(n) = eval(['sub2ind(ngrid.*ones(1,ndim)' idnames ');']);
end
% Quality based on the number of particles in each hypercube
REP.quality = zeros(ngrid,2);
ids = unique(REP.grid_idx);
for i = 1:length(ids)
REP.quality(i,1) = ids(i); % First, the hypercube's identifier
REP.quality(i,2) = 10/sum(REP.grid_idx==ids(i)); % Next, its quality
end
end
% Function that selects the leader performing a roulette wheel selection
% based on the quality of each hypercube
function selected = selectLeader(REP)
% Roulette wheel
prob = cumsum(REP.quality(:,2)); % Cumulated probs
sel_hyp = REP.quality(find(rand(1,1)*max(prob)<=prob,1,'first'),1); % Selected hypercube
% Select the index leader as a random selection inside that hypercube
idx = 1:1:length(REP.grid_idx);
selected = idx(REP.grid_idx==sel_hyp);
selected = selected(randi(length(selected)));
end
% Function that deletes an excess of particles inside the repository using
% crowding distances
function REP = deleteFromRepository(REP,n_extra,ngrid)
% Compute the crowding distances
crowding = zeros(size(REP.pos,1),1);
for m = 1:1:size(REP.pos_fit,2)
[m_fit,idx] = sort(REP.pos_fit(:,m),'ascend');
m_up = [m_fit(2:end); Inf];
m_down = [Inf; m_fit(1:end-1)];
distance = (m_up-m_down)./(max(m_fit)-min(m_fit));
[~,idx] = sort(idx,'ascend');
crowding = crowding + distance(idx);
end
crowding(isnan(crowding)) = Inf;
% Delete the extra particles with the smallest crowding distances
[~,del_idx] = sort(crowding,'ascend');
del_idx = del_idx(1:n_extra);
REP.pos(del_idx,:) = [];
REP.pos_fit(del_idx,:) = [];
REP = updateGrid(REP,ngrid);
end
% Function that performs the mutation of the particles depending on the
% current generation
function POS = mutation(POS,gen,maxgen,Np,var_max,var_min,nVar,u_mut)
% Sub-divide the swarm in three parts [2]
fract = Np/3 - floor(Np/3);
if(fract<0.5), sub_sizes =[ceil(Np/3) round(Np/3) round(Np/3)];
else sub_sizes =[round(Np/3) round(Np/3) floor(Np/3)];
end
cum_sizes = cumsum(sub_sizes);
% First part: no mutation
% Second part: uniform mutation
nmut = round(u_mut*sub_sizes(2));
if(nmut>0)
idx = cum_sizes(1) + randperm(sub_sizes(2),nmut);
POS(idx,:) = repmat((var_max-var_min)',nmut,1).*rand(nmut,nVar) + repmat(var_min',nmut,1);
end
% Third part: non-uniform mutation
per_mut = (1-gen/maxgen)^(5*nVar); % Percentage of mutation
nmut = round(per_mut*sub_sizes(3));
if(nmut>0)
idx = cum_sizes(2) + randperm(sub_sizes(3),nmut);
POS(idx,:) = repmat((var_max-var_min)',nmut,1).*rand(nmut,nVar) + repmat(var_min',nmut,1);
end
end