-
Notifications
You must be signed in to change notification settings - Fork 1
/
demoInit.m
254 lines (233 loc) · 10.3 KB
/
demoInit.m
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
clc; clearvars; close all; rng(0);
nRepeats=8;
nFs=1:8; % number of features
nMFs=2; % number of MFs in each input domain, used in MBGD_RDA and MBGD_RDA_T
nRs=nMFs.^nFs; % number of rules
alphas=0.01; % initial learning rate
lambda=0.05; % L2 regularization coefficient
Ps=0.5; % DropRule Rate
gammaPs=0.5; % powerball param
nIt=1000; % number of iterations
Nbs=64; % batch size
LN0={'rand-RDpA','kM-RDpA','FCM-RDpA'};
LN=cell(1,length(LN0)*length(nRs)+1);
LN(1)={'RR'};
for i=1:length(nRs)
LN(2+(i-1)*length(LN0):1+i*length(LN0))=strcat(LN0, ['-nR' num2str(nRs(i))]);
end
nAlgs=length(LN);
datasets={'Concrete-CS';'Concrete-Flow';'Concrete-Slump';'tecator-fat';'tecator-moisture';'tecator-protein';'Yacht';'autoMPG';'NO2';'PM10';'Housing';'CPS';'EnergyEfficiency-Cooling';'EnergyEfficiency-Heating';'Concrete';'Airfoil';'Wine-red';'Abalone';'Abalone-onehot';'Wine-white';'PowerPlant';'Protein'};
datasets=datasets(1)
% Display results in parallel computing
dqWorker = parallel.pool.DataQueue; afterEach(dqWorker, @(data) fprintf('%d-%d ', data{1},data{2})); % print progress of parfor
[RMSEtrain,RMSEtest,RMSEtune]=deal(cellfun(@(u)nan(length(datasets),nAlgs,nIt),cell(nRepeats,1),'UniformOutput',false));
[InitTimes,times,BestP,Bestalpha,BestgammaP]=deal(cellfun(@(u)nan(length(datasets),nAlgs),cell(nRepeats,1),'UniformOutput',false));
BestmIter=cellfun(@(u)ones(length(datasets),nAlgs),cell(nRepeats,1),'UniformOutput',false);
thres=cellfun(@(u)inf(length(datasets),nAlgs),cell(nRepeats,1),'UniformOutput',false);
delete(gcp('nocreate'))
parpool(nRepeats);
parfor r=1:nRepeats
dataDisp=cell(1,2); dataDisp{1}=r;
for s=1:length(datasets)
dataDisp{2} = s; send(dqWorker,dataDisp); % Display progress in parfor
temp=load(['./' datasets{s} '.mat']);
data=temp.data;
X=data(:,1:end-1); y=data(:,end); y=y-mean(y);
X = zscore(X); [N0,M]=size(X);
N=round(N0*.7);
idsTrain=datasample(1:N0,N,'replace',false);
XTrain=X(idsTrain,:); yTrain=y(idsTrain);
XTest=X; XTest(idsTrain,:)=[];
yTest=y; yTest(idsTrain)=[];
% validation data
N1=round(N0*.15);
idsTune=datasample(1:(N0-N),N1,'replace',false);
XTune=XTest(idsTune,:); yTune=yTest(idsTune);
XTest(idsTune,:)=[]; yTest(idsTune)=[];
idsTest=1:N0;idsTest([idsTrain idsTune])=[];
trainInd=idsTrain;
testInd=1:N0;testInd(idsTrain)=[];
valInd=testInd(idsTune);
testInd(idsTune)=[];
MXTrain=mean(XTrain);
XTrain=XTrain-MXTrain; XTune=XTune-MXTrain; XTest=XTest-MXTrain;
nFs0=nFs;
nFs0(nFs0>M)=[];
nFs0(nMFs.^nFs0>N)=[];
% disp(nFs0)
%% 1. Ridge regression
tic
id=1;
b = ridge(yTrain,XTrain,lambda,0);
RMSEtrain{r}(s,id,:) = sqrt(mean((yTrain-[ones(N,1) XTrain]*b).^2));
RMSEtune{r}(s,id,:) = sqrt(mean((yTune-[ones(N1,1) XTune]*b).^2));
RMSEtest{r}(s,id,:) = sqrt(mean((yTest-[ones(length(yTest),1) XTest]*b).^2));
times{r}(s,id)=toc;
for nFeatures=nFs0
nRules=nMFs^nFeatures;
[WPCA,XTrainP]=pca(XTrain,'NumComponents',nFeatures);
XTuneP=XTune*WPCA; XTestP=XTest*WPCA;
WA=[zeros(1,size(WPCA,2)); WPCA];
%% random init
tic
rC3= 2*rand(nRules,M)-1;
rSigma3= rand(nRules,M)*5;
rW3= 2*rand(nRules,M+1)-1;
rt3=toc;
%% rand-RDpA
tic;
id=id+1;
InitTimes{r}(s,id)=rt3;
for P=Ps
for alpha=alphas
for gammaP=gammaPs
[tmp,tmpt]=FCM_RDpA(XTrain,yTrain,{XTune,XTest},{yTune,yTest},alpha,lambda,P,gammaP,nRules,nIt,Nbs,rC3,rSigma3,rW3);
if min(tmpt{1})<thres{r}(s,id)||~isfinite(thres{r}(s,id))
[thres{r}(s,id),BestmIter{r}(s,id)]=min(tmpt{1});
BestP{r}(s,id)=P;
Bestalpha{r}(s,id)=alpha;
BestgammaP{r}(s,id)=gammaP;
[RMSEtrain{r}(s,id,:),RMSEtune{r}(s,id,:),RMSEtest{r}(s,id,:)]=deal(tmp,tmpt{1},tmpt{2});
end
end
end
end
times{r}(s,id)=toc;
%% k-Means
tic
kW3=zeros(nRules,M+1); % Rule consequents
[ids,kC3] = kmeans(XTrain,nRules,'replicate',3);
kSigma3=kC3;
for ir=1:nRules
kSigma3(ir,:)=std(XTrain(ids==ir,:));
kW3(ir,1)=mean(yTrain(ids==ir));
end
kSigma3(kSigma3==0)=mean(kSigma3(:));
kt3=toc;
%% kM-RDpA
tic;
id=id+1;
InitTimes{r}(s,id)=kt3;
for P=Ps
for alpha=alphas
for gammaP=gammaPs
[tmp,tmpt]=FCM_RDpA(XTrain,yTrain,{XTune,XTest},{yTune,yTest},alpha,lambda,P,gammaP,nRules,nIt,Nbs,kC3,kSigma3,kW3);
if min(tmpt{1})<thres{r}(s,id)||~isfinite(thres{r}(s,id))
[thres{r}(s,id),BestmIter{r}(s,id)]=min(tmpt{1});
BestP{r}(s,id)=P;
Bestalpha{r}(s,id)=alpha;
BestgammaP{r}(s,id)=gammaP;
[RMSEtrain{r}(s,id,:),RMSEtune{r}(s,id,:),RMSEtest{r}(s,id,:)]=deal(tmp,tmpt{1},tmpt{2});
end
end
end
end
times{r}(s,id)=toc;
%% Fuzzy C-Means (FCM)
tic
W3=zeros(nRules,M+1); % Rule consequents
% FCM initialization
[C3,U] = FuzzyCMeans(XTrain,nRules,[2 100 0.001 0]);
Sigma3=C3;
for ir=1:nRules
Sigma3(ir,:)=std(XTrain,U(ir,:));
W3(ir,1)=U(ir,:)*yTrain/sum(U(ir,:));
end
Sigma3(Sigma3==0)=mean(Sigma3(:));
t3=toc;
%% FCM-RDpA
tic;
id=id+1;
InitTimes{r}(s,id)=t3;
for P=Ps
for alpha=alphas
for gammaP=gammaPs
[tmp,tmpt]=FCM_RDpA(XTrain,yTrain,{XTune,XTest},{yTune,yTest},alpha,lambda,P,gammaP,nRules,nIt,Nbs,C3,Sigma3,W3);
if min(tmpt{1})<thres{r}(s,id)||~isfinite(thres{r}(s,id))
[thres{r}(s,id),BestmIter{r}(s,id)]=min(tmpt{1});
BestP{r}(s,id)=P;
Bestalpha{r}(s,id)=alpha;
BestgammaP{r}(s,id)=gammaP;
[RMSEtrain{r}(s,id,:),RMSEtune{r}(s,id,:),RMSEtest{r}(s,id,:)]=deal(tmp,tmpt{1},tmpt{2});
end
end
end
end
times{r}(s,id)=toc;
end
end
end
save('demoInit.mat','RMSEtrain','RMSEtune','RMSEtest','times','InitTimes','BestP','Bestalpha','BestmIter','BestgammaP','datasets','nAlgs','Nbs','LN','LN0','lambda','nRepeats','nFs','nMFs','alphas','Ps','gammaPs','thres','nRs','nIt');
%% Plot results
ids=1:length(LN);
[tmp,ttmp]=deal(nan(length(datasets),length(LN),nRepeats));
for s=1:length(datasets)
ttmp0=cellfun(@(u)squeeze(u(s,ids)),times,'UniformOutput',false);
ttmp(s,ids,:)=cat(1,ttmp0{:})';
for id=1:length(LN)
tmp(s,id,:)=cell2mat(cellfun(@(u,m)squeeze(u(s,id,m(s,id))),RMSEtest,BestmIter,'UniformOutput',false));
end
end
A=[nanmean(nanmean(tmp(:,ids,:),1),3);
nanstd(nanmean(tmp(:,ids,:),1),[],3);
nanmean(nanmean(ttmp(:,ids,:),1),3);
nanstd(nanmean(ttmp(:,ids,:),1),[],3);
nanmean(cat(1,Bestalpha{:}),1);
nanmean(cat(1,BestP{:}),1);
nanmean(cat(1,BestmIter{:}),1)
nanmean(cat(1,thres{:}),1)];
a=squeeze(nanmean(tmp(:,ids,:),3));
a=[a;nanmean(a,1)]; sa=sort(a,2);
b=a==sa(:,1);c=a==sa(:,2);
at=squeeze(nanmean(ttmp(:,ids,:),3));
aa=nanmean(cat(3,Bestalpha{:}),3); aa=[aa;nanmean(aa,1)];
ap=nanmean(cat(3,BestP{:}),3); ap=[ap;nanmean(ap,1)];
am=nanmean(cat(3,BestmIter{:}),3); am=[am;nanmean(am,1)];
avgRMSE=nanmean(nanmean(tmp,1),3);
stdRMSE=nanstd(nanmean(tmp,1),[],3);
[iavgRMSE,istdRMSE,iavgTIME,istdTIME]=deal(nan(length(LN0)+1,length(nRs)));
iavgRMSE(1,:)=repmat(avgRMSE(:,1),1,length(nRs));
istdRMSE(1,:)=repmat(stdRMSE(:,1),1,length(nRs));
for i=2:length(LN0)+1
iavgRMSE(i,:)=avgRMSE(:,i:length(LN0):end);
istdRMSE(i,:)=stdRMSE(:,i:length(LN0):end);
end
avgTIME=nanmean(nanmean(ttmp,1),3);
stdTIME=nanstd(nanmean(ttmp,1),[],3);
iavgTIME(1,:)=repmat(avgTIME(:,1),1,length(nRs));
istdTIME(1,:)=repmat(stdTIME(:,1),1,length(nRs));
for i=2:length(LN0)+1
iavgTIME(i,:)=avgTIME(:,i:length(LN0):end);
istdTIME(i,:)=stdTIME(:,i:length(LN0):end);
end
close all
color={'k','g','b','r','m','c','#0072BD','#D95319','#EDB120','#7E2F8E','#77AC30','#4DBEEE','#A2142F'};
style={'-','--'};
lineStyles=cell(2,length(color)*length(style));
for i=1:length(color)
for j=1:length(style)
lineStyles{1,length(style)*(i-1)+j}=color{i};
lineStyles{2,length(style)*(i-1)+j}=style{j};
end
end
figure;
set(gcf,'DefaulttextFontName','times new roman','DefaultaxesFontName','times new roman','defaultaxesfontsize',12);
idM=1:length(LN0);
f=mat2cell(permute(tmp(:,idM+(1:length(LN0):size(tmp,2)-length(LN0))',:),[3,2,1]),ones(1,nRepeats));
f=cellfun(@(x)permute(x,[2,3,1]),f,'UniformOutput',false);
RR=mat2cell(permute(tmp(:,1,:),[3,2,1]),ones(1,nRepeats));
RR=cellfun(@(x)permute(x,[2,3,1]),RR,'UniformOutput',false);
fR=cellfun(@(x,y)reshape(nanmean(x./y,2),length(nRs),[])',f,RR,'UniformOutput',false);
fR=cat(3,fR{:});
savgRMSE=nanmean(fR,3);
sstdRMSE=nanstd(fR,[],3);
for i=1:length(idM)
errorbar(nRs, savgRMSE(i,:),sstdRMSE(i,:),'Color',lineStyles{1,i},'LineStyle',lineStyles{2,i},'linewidth',2);
hold on;
end
set(gca,'XTick',nRs);
xlabel('$R$','interpreter','latex');
ylabel('Average normalized test RMSE');
set(gca,'yscale','log','xscale','log');
box on; axis tight;
legend(LN0(idM),'NumColumns',1,'Location','eastoutside');