Skip to content

Commit

Permalink
add new algorithms, problems and metrics
Browse files Browse the repository at this point in the history
  • Loading branch information
@anonymone
anonymone committed Apr 14, 2019
1 parent fd90fdc commit ccd0f59
Show file tree
Hide file tree
Showing 13 changed files with 687 additions and 0 deletions.
22 changes: 22 additions & 0 deletions PlatEMO/Algorithms/MOEA-D-FRRMAB/CreditAssignment.m
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,22 @@
function FRR = CreditAssignment(SW,D)
% Credit assignment

%------------------------------- Copyright --------------------------------
% Copyright (c) 2018-2019 BIMK Group. You are free to use the PlatEMO for
% research purposes. All publications which use this platform or any code
% in the platform should acknowledge the use of "PlatEMO" and reference "Ye
% Tian, Ran Cheng, Xingyi Zhang, and Yaochu Jin, PlatEMO: A MATLAB platform
% for evolutionary multi-objective optimization [educational forum], IEEE
% Computational Intelligence Magazine, 2017, 12(4): 73-87".
%--------------------------------------------------------------------------

K = 4; % Number of operators
Reward = zeros(1,K);
for i = 1 : K
Reward(i) = sum(SW(2,SW(1,:)==i));
end
[~,Rank] = sort(Reward,'descend');
[~,Rank] = sort(Rank);
Decay = D.^Rank.*Reward;
FRR = Decay./sum(Decay);
end
19 changes: 19 additions & 0 deletions PlatEMO/Algorithms/MOEA-D-FRRMAB/FRRMAB.m
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,19 @@
function op = FRRMAB(FRR,SW,C)
% Bandit-based operator selection

%------------------------------- Copyright --------------------------------
% Copyright (c) 2018-2019 BIMK Group. You are free to use the PlatEMO for
% research purposes. All publications which use this platform or any code
% in the platform should acknowledge the use of "PlatEMO" and reference "Ye
% Tian, Ran Cheng, Xingyi Zhang, and Yaochu Jin, PlatEMO: A MATLAB platform
% for evolutionary multi-objective optimization [educational forum], IEEE
% Computational Intelligence Magazine, 2017, 12(4): 73-87".
%--------------------------------------------------------------------------

if any(FRR==0) || any(SW(1,:)==0)
op = randi(length(FRR));
else
n = hist(SW(1,:),1:length(FRR));
[~,op] = max(FRR+C*sqrt(2*log(sum(n))./n));
end
end
50 changes: 50 additions & 0 deletions PlatEMO/Algorithms/MOEA-D-FRRMAB/FourDE.m
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,50 @@
function Offspring = FourDE(op,x,x1,x2,x3,x4,x5)
% Four different DE operators

%------------------------------- Copyright --------------------------------
% Copyright (c) 2018-2019 BIMK Group. You are free to use the PlatEMO for
% research purposes. All publications which use this platform or any code
% in the platform should acknowledge the use of "PlatEMO" and reference "Ye
% Tian, Ran Cheng, Xingyi Zhang, and Yaochu Jin, PlatEMO: A MATLAB platform
% for evolutionary multi-objective optimization [educational forum], IEEE
% Computational Intelligence Magazine, 2017, 12(4): 73-87".
%--------------------------------------------------------------------------

%% Parameter setting
[CR,F,proM,disM,K] = deal(1,0.5,1,20,0.5);
D = length(x.dec);
Global = GLOBAL.GetObj();

%% Differental evolution
switch op
case 1
% DE/rand/1
v = x.dec + F*(x1.dec-x2.dec);
case 2
% DE/rand/2
v = x.dec + F*(x1.dec-x2.dec) + F*(x3.dec-x4.dec);
case 3
% DE/current-to-rand/2
v = x.dec + K*(x.dec-x1.dec) + F*(x2.dec-x3.dec) + F*(x4.dec-x5.dec);
case 4
% DE/current-to-rand/1
v = x.dec + K*(x.dec-x1.dec) + F*(x2.dec-x3.dec);
end
Offspring = x.dec;
Site = rand(1,D) < (CR+(op>2));
Offspring(Site) = v(Site);

%% Polynomial mutation
Lower = Global.lower;
Upper = Global.upper;
Site = rand(1,D) < proM/D;
mu = rand(1,D);
temp = Site & mu<=0.5;
Offspring = min(max(Offspring,Lower),Upper);
Offspring(temp) = Offspring(temp)+(Upper(temp)-Lower(temp)).*((2.*mu(temp)+(1-2.*mu(temp)).*...
(1-(Offspring(temp)-Lower(temp))./(Upper(temp)-Lower(temp))).^(disM+1)).^(1/(disM+1))-1);
temp = Site & mu>0.5;
Offspring(temp) = Offspring(temp)+(Upper(temp)-Lower(temp)).*(1-(2.*(1-mu(temp))+2.*(mu(temp)-0.5).*...
(1-(Upper(temp)-Offspring(temp))./(Upper(temp)-Lower(temp))).^(disM+1)).^(1/(disM+1)));
Offspring = INDIVIDUAL(Offspring);
end
92 changes: 92 additions & 0 deletions PlatEMO/Algorithms/MOEA-D-FRRMAB/MOEADFRRMAB.m
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,92 @@
function MOEADFRRMAB(Global)
% <algorithm> <M>
% MOEA/D with fitness-rate-rank-based multiarmed bandit
% C --- 5 --- Scaling factor in bandit-based operator selection
% W --- --- Size of sliding window
% D --- 1 --- Decaying factor in calculating credit value

%------------------------------- Reference --------------------------------
% K. Li, A. Fialho, S. Kwong, and Q. Zhang, Adaptive operator selection
% with bandits for a multiobjective evolutionary algorithm based on
% decomposition, IEEE Transactions on Evolutionary Computation, 2014,
% 18(1): 114-130.
%------------------------------- Copyright --------------------------------
% Copyright (c) 2018-2019 BIMK Group. You are free to use the PlatEMO for
% research purposes. All publications which use this platform or any code
% in the platform should acknowledge the use of "PlatEMO" and reference "Ye
% Tian, Ran Cheng, Xingyi Zhang, and Yaochu Jin, PlatEMO: A MATLAB platform
% for evolutionary multi-objective optimization [educational forum], IEEE
% Computational Intelligence Magazine, 2017, 12(4): 73-87".
%--------------------------------------------------------------------------

%% Parameter setting
[C,W,D] = Global.ParameterSet(5,ceil(Global.N/2),1);

%% Generate the weight vectors
[Weight,Global.N] = UniformPoint(Global.N,Global.M);
% Size of neighborhood
T = 20;
% Maximum number of solutions replaced by each offspring
nr = 2;

%% Detect the neighbours of each solution
B = pdist2(Weight,Weight);
[~,B] = sort(B,2);
B = B(:,1:T);

%% Generate random population
Population = Global.Initialization();
Z = min(Population.objs,[],1);
% Utility for each subproblem
Pi = ones(Global.N,1);
% Old Tchebycheff function value of each solution on its subproblem
oldObj = max(abs((Population.objs-repmat(Z,Global.N,1)).*Weight),[],2);

%% Optimization
FRR = zeros(1,4); % Credit value of each operator
SW = zeros(2,W); % Sliding window
while Global.NotTermination(Population)
for subgeneration = 1 : 5
% Choose I
Bounday = find(sum(Weight<1e-3,2)==Global.M-1)';
I = [Bounday,TournamentSelection(10,floor(Global.N/5)-length(Bounday),-Pi)];

% For each solution in I
for i = I
% Bandit-based operator selection
op = FRRMAB(FRR,SW,C);

% Choose the parents
if rand < 0.9
P = B(i,randperm(end));
else
P = randperm(Global.N);
end

% Generate an offspring
Offspring = FourDE(op,Population(i),Population(P(1)),Population(P(2)),Population(P(3)),Population(P(4)),Population(P(5)));

% Update the ideal point
Z = min(Z,Offspring.obj);

% Update the solutions in P by Tchebycheff approach
g_old = max(abs(Population(P).objs-repmat(Z,length(P),1)).*Weight(P,:),[],2);
g_new = max(repmat(abs(Offspring.obj-Z),length(P),1).*Weight(P,:),[],2);
replace = find(g_old>=g_new,nr);
Population(P(replace)) = Offspring;
FIR = sum((g_old(replace)-g_new(replace))./g_old(replace));
SW = [SW(1,2:end),op;SW(2,2:end),FIR];
FRR = CreditAssignment(SW,D);
end
end
if ~mod(Global.gen,10)
% Update Pi for each solution
newObj = max(abs((Population.objs-repmat(Z,Global.N,1)).*Weight),[],2);
DELTA = (oldObj-newObj)./oldObj;
Temp = DELTA < 0.001;
Pi(~Temp) = 1;
Pi(Temp) = (0.95+0.05*DELTA(Temp)/0.001).*Pi(Temp);
oldObj = newObj;
end
end
end
73 changes: 73 additions & 0 deletions PlatEMO/Metrics/CPF.m
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,73 @@
function Score = CPF(PopObj,PF)
% <metric> <max>
% Coverage of Pareto front

%------------------------------- Copyright --------------------------------
% Copyright (c) 2018-2019 BIMK Group. You are free to use the PlatEMO for
% research purposes. All publications which use this platform or any code
% in the platform should acknowledge the use of "PlatEMO" and reference "Ye
% Tian, Ran Cheng, Xingyi Zhang, and Yaochu Jin, PlatEMO: A MATLAB platform
% for evolutionary multi-objective optimization [educational forum], IEEE
% Computational Intelligence Magazine, 2017, 12(4): 73-87".
%--------------------------------------------------------------------------

if size(PF,1) > 1
%% Normalization
fmin = min(PF,[],1);
fmax = max(PF,[],1);
PopObj = (PopObj-repmat(fmin,size(PopObj,1),1))./repmat(fmax-fmin,size(PopObj,1),1);
PF = (PF-repmat(fmin,size(PF,1),1))./repmat(fmax-fmin,size(PF,1),1);

%% Map to the Pareto front
[~,Close] = min(pdist2(PopObj,PF),[],2);
PopObj = PF(Close,:);

%% Calculate the indicator value
VPF = Coverage(map(PF,PF),inf);
V = Coverage(map(PopObj,PF),VPF/size(PopObj,1));
Score = V./VPF;
else
fmin = min(PopObj,[],1);
fmax = max(PopObj,[],1);
PopObj = (PopObj-repmat(fmin,size(PopObj,1),1))./repmat(fmax-fmin,size(PopObj,1),1);
Score = Coverage(map(PopObj,PopObj),1/size(PopObj,1));
end
end

function y = map(x,PF)
% Project the points in an (M-1)-d manifold to an (M-1)-d unit hypercube

[N,M] = size(x);
x = x - repmat((sum(x,2)-1)/M,1,M);
PF = PF - repmat((sum(PF,2)-1)/M,1,M);
x = x - min(PF);
x = x./repmat(sum(x,2),1,M);
x = max(1e-6,x);
y = zeros(N,M-1);
for i = 1 : N
c = ones(1,M);
k = find(x(i,:)~=0,1);
for j = k+1 : M
temp = x(i,j)/x(i,k)*prod(c(M-j+2:M-k));
c(M-j+1) = 1/(temp+1);
end
y(i,:) = c(1:M-1);
end
y = y.^repmat(M-1:-1:1,N,1);
end

function V = Coverage(P,maxv)
% Calculate the hypervolume of each point's monopolized hypercube

[N,M] = size(P);
L = zeros(N,1);
for x = 1 : N
P1 = P;
P1(x,:) = inf;
L(x) = min(max(abs(P1-repmat(P(x,:),N,1)),[],2));
end
L = min(L,maxv.^(1/M));
Lower = max(0,P-repmat(L/2,1,M));
Upper = min(1,P+repmat(L/2,1,M));
V = sum(prod(Upper-Lower,2));
end
46 changes: 46 additions & 0 deletions PlatEMO/Problems/IMOP/IMOP1.m
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,46 @@
classdef IMOP1 < PROBLEM
% <problem> <IMOP>
% Benchmark MOP with irregular Pareto front
% a1 --- 0.05 --- Parameter a1
% K --- 5 --- Parameter K

%------------------------------- Copyright --------------------------------
% Copyright (c) 2018-2019 BIMK Group. You are free to use the PlatEMO for
% research purposes. All publications which use this platform or any code
% in the platform should acknowledge the use of "PlatEMO" and reference "Ye
% Tian, Ran Cheng, Xingyi Zhang, and Yaochu Jin, PlatEMO: A MATLAB platform
% for evolutionary multi-objective optimization [educational forum], IEEE
% Computational Intelligence Magazine, 2017, 12(4): 73-87".
%--------------------------------------------------------------------------

properties(Access = private)
a1 = 0.05; % Parameter a1
K = 5; % Parameter K
end
methods
%% Initialization
function obj = IMOP1()
[obj.a1,obj.K] = obj.Global.ParameterSet(0.05,5);
obj.Global.M = 2;
if isempty(obj.Global.D)
obj.Global.D = 10;
end
obj.Global.lower = zeros(1,obj.Global.D);
obj.Global.upper = ones(1,obj.Global.D);
obj.Global.encoding = 'real';
end
%% Calculate objective values
function PopObj = CalObj(obj,PopDec)
y1 = mean(PopDec(:,1:obj.K),2).^obj.a1;
g = sum((PopDec(:,obj.K+1:end)-0.5).^2,2);
PopObj(:,1) = g + cos(y1*pi/2).^8;
PopObj(:,2) = g + sin(y1*pi/2).^8;
end
%% Sample reference points on Pareto front
function P = PF(obj,N)
x = linspace(0.5^4,1,floor(N/2))';
P(:,1) = [x;(1-x.^0.25).^4];
P(:,2) = [(1-x.^0.25).^4;x];
end
end
end
46 changes: 46 additions & 0 deletions PlatEMO/Problems/IMOP/IMOP2.m
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,46 @@
classdef IMOP2 < PROBLEM
% <problem> <IMOP>
% Benchmark MOP with irregular Pareto front
% a1 --- 0.05 --- Parameter a1
% K --- 5 --- Parameter K

%------------------------------- Copyright --------------------------------
% Copyright (c) 2018-2019 BIMK Group. You are free to use the PlatEMO for
% research purposes. All publications which use this platform or any code
% in the platform should acknowledge the use of "PlatEMO" and reference "Ye
% Tian, Ran Cheng, Xingyi Zhang, and Yaochu Jin, PlatEMO: A MATLAB platform
% for evolutionary multi-objective optimization [educational forum], IEEE
% Computational Intelligence Magazine, 2017, 12(4): 73-87".
%--------------------------------------------------------------------------

properties(Access = private)
a1 = 0.05; % Parameter a1
K = 5; % Parameter K
end
methods
%% Initialization
function obj = IMOP2()
[obj.a1,obj.K] = obj.Global.ParameterSet(0.05,5);
obj.Global.M = 2;
if isempty(obj.Global.D)
obj.Global.D = 10;
end
obj.Global.lower = zeros(1,obj.Global.D);
obj.Global.upper = ones(1,obj.Global.D);
obj.Global.encoding = 'real';
end
%% Calculate objective values
function PopObj = CalObj(obj,PopDec)
y1 = mean(PopDec(:,1:obj.K),2).^obj.a1;
g = sum((PopDec(:,obj.K+1:end)-0.5).^2,2);
PopObj(:,1) = g + cos(y1*pi/2).^0.5;
PopObj(:,2) = g + sin(y1*pi/2).^0.5;
end
%% Sample reference points on Pareto front
function P = PF(obj,N)
x = linspace(0,0.5^0.25,floor(N/2))';
P(:,1) = [x;(1-x.^4).^0.25];
P(:,2) = [(1-x.^4).^0.25;x];
end
end
end
Loading

0 comments on commit ccd0f59

Please sign in to comment.