-
Notifications
You must be signed in to change notification settings - Fork 5
/
Copy pathwsabiM.m
382 lines (289 loc) · 15.2 KB
/
wsabiM.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
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
function [ log_mu, log_Var, clktime, xxIter, hyp ] = wsabiM( ...
range, ... 1) 2 x D matrix, lower bnd top row.
priorMu, ... 2) Gaussian prior mean, D x 1.
priorVar, ... 3) Gaussian prior covariance, D x D.
kernelVar, ... 4) Initial input length scales, D x D.
lambda, ... 5) Initial output length scale.
alpha, ... 6) Alpha offset fraction, as in paper.
numSamples, ... 7) Number of BBQ samples to run.
loglikhandle, ... 8) Handle to log-likelihood function.
printing ) ... 9) If true, print intermediate output.
% Output structures:
% log_mu: log of the integral posterior mean.
% log_var: log of the integral posterior variance.
% clktime: vector of times per iteration, may want to cumulative sum.
% xxIter: numSamples x D array of sample locations used to build model.
% hyp: integral hyperparameters.
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% Relabel prior mean and covariance for brevity of code.
bb = priorMu;
BB = diag(priorVar)';
% Relabel input hyperparameters for brevity of code.
VV = diag(kernelVar)';
jitterNoise = 1e-6; % Jitter on the GP model over the log likelihood.
numEigs = inf; % If trying to use Nystrom ** NOT RECOMMENDED **
hypOptEvery = 1; % 1 => optimise hyperparameters every iteration.
dim = length(bb); % Dimensionality of integral.
% Limit absolute range of likelihood model hyperparameters for stability.
hypLims = 30*ones(1,dim+1);
% Allocate Storage
mu = zeros(numSamples-1,1);
logscaling = zeros(numSamples-1,1);
Var = zeros(numSamples-1,1);
clktime = zeros(numSamples-1,1);
lHatD_0_tmp = zeros(numSamples,1);
loglHatD_0_tmp = zeros(size(lHatD_0_tmp));
hyp = zeros(1,1+dim);
% Minimiser options (fmincon for hyperparameters)
options1 = optimset('fmincon');
options1.Display = 'none';
options1.GradObj = 'off';
options1.Algorithm = 'active-set';
options1.TolX = 1e-5;
options1.TolFun = 1e-5;
options1.MaxTime = 0.5;
options1.MaxFunEvals = 50;
%options1.UseParallel = 'always';
options1.AlwaysHonorConstraints = 'true';
% Minimiser options (fmincon if desired for active sampling)
options2 = optimset('fmincon');
options2.Display = 'none';
options2.GradObj = 'on';
%options2.DerivativeCheck = 'on';
options2.TolX = 1e-5;
options2.TolFun = 1e-5;
%options2.MaxTime = 0.5;
%options2.MaxFunEvals = 75;
options2.UseParallel = 'always';
options2.AlwaysHonorConstraints = 'true';
% Minimiser options (CMAES - advised for active sampling)
opts = cmaes('defaults');
opts.LBounds = range(1,:)';
opts.UBounds = range(2,:)';
opts.DispModulo = Inf;
opts.DispFinal = 'off';
opts.SaveVariables = 'off';
%opts.EvalParallel = 'on';
%opts.PopSize = 100;
%opts.Restarts = 1;
% Initial Sample set to priorMean:
xx = zeros(numSamples,dim);
xx(end,:) = bb+1e-6;
currNumSamples = 1;
if printing
fprintf('Iter: ');
end
for t = 1:numSamples - 1
if printing
if ~mod(t,100)
prstr = sprintf('Log Current Mean Integral: %g', ...
mu(t-1)*exp(logscaling(t-1)));
fprintf(prstr);
pause(1);
fprintf(repmat('\b',1,length(prstr)))
end
if t > 1
fprintf(repmat('\b',1,length(num2str(t-1))));
end
fprintf('%i',t);
end
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% Pre-process new samples -- i.e. convert to log space etc.
% Get batch of samples & variables from stack.
tmpT = cputime;
xxIter = xx(numSamples-currNumSamples+1:end,:); % Curr samples
% Call loglik handle for latest sample.
loglHatD_0_tmp(numSamples-currNumSamples+1) = ...
loglikhandle( xxIter(1,:) );
% Find the max in log space.
logscaling(t) = max(loglHatD_0_tmp(numSamples-currNumSamples+1:end));
% Scale batch by max, and exponentiate.
lHatD_0_tmp(numSamples-currNumSamples+1:end) = ...
exp(loglHatD_0_tmp(numSamples-currNumSamples+1:end) - logscaling(t));
% Evaluate the offset, alpha fraction of minimum value seen.
aa = alpha * min( lHatD_0_tmp(numSamples-currNumSamples+1:end) );
% Transform into sqrt space.
lHatD = sqrt(abs(lHatD_0_tmp(numSamples-currNumSamples+1:end)- aa)*2);
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% ML-II On GP Likelihood model hyperparameters
hyp(1) = log( lambda );
hyp(2:end) = log(VV);
if currNumSamples > 3 && ~mod(currNumSamples,hypOptEvery)
if currNumSamples < numEigs + 1
hypLogLik = @(x) logLikGPDim(xxIter, lHatD, x);
else
hypLogLik = @(x) logLikGPDimNystrom(xxIter, lHatD, x, numEigs);
end
[hyp] = fmincon(hypLogLik, ...
hyp,[],[],[],[],-hypLims,hypLims,[],options1);
end
lambda = exp(hyp(1));
VV = exp(hyp(2:end));
% Scale samples by input length scales.
xxIterScaled = xxIter .* repmat(sqrt(1./VV),currNumSamples,1);
% Squared distance matrix
dist2 = pdist2_squared_fast(xxIterScaled, xxIterScaled);
% Evaluate Gram matrix
Kxx = lambda.^2 * (1/(prod(2*pi*VV).^0.5)) * exp(-0.5*dist2);
Kxx = Kxx + ...
lambda.^2*(1/(prod(2*pi*VV).^0.5))*jitterNoise*eye(size(Kxx));
Kxx = Kxx/2 + Kxx'/2; % Make sure symmetric for stability.
% Invert Gram matrix.
if currNumSamples < numEigs + 1
invKxx = Kxx \ eye(size(Kxx));
else % If using nystrom
idx = randperm( length(xxIter(:,1)) );
xxuScaled = xxIterScaled( idx(1:numEigs), : );
xxsScaled = xxIterScaled;
AA1 = pdist2_squared_fast(xxuScaled,xxuScaled);
AA2 = pdist2_squared_fast(xxsScaled,xxuScaled);
Kuu = lambda^2 * 1/sqrt(det(2*pi*VV)) * (exp( -0.5 * AA1 ) + ...
jitterNoise*eye(size(AA1)));
Ksu = lambda^2 * 1/sqrt(det(2*pi*VV)) * exp( -0.5 * AA2 );
[eVec, eVal] = eig(Kuu);
eVec = Ksu * ...
repmat(sqrt(numEigs)./diag(eVal)',length(eVal(:,1)),1) ...
.* eVec;
eVal = eVal / numEigs;
Z = jitterNoise * diag(1./diag(eVal)) + eVec'*eVec;
invKxx = (1./jitterNoise)*(eye(currNumSamples)-eVec*(Z \ eVec'));
end
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% Expected value of integral:
ww = invKxx * lHatD;
Yvar = (VV.*VV + 2*VV.*BB)./VV;
postProd = ww*ww';
xx2sq = xxIter .* repmat(sqrt(1./Yvar),currNumSamples,1);
bbch = bb .* sqrt(1./Yvar);
xx2sqFin = pdist2_squared_fast(xx2sq,bbch);
xxIterScaled2 = xxIter .* repmat(sqrt(BB./(VV.*Yvar)),currNumSamples,1);
dist4 = pdist2_squared_fast(xxIterScaled2,xxIterScaled2);
% Sigma^2 term:
sig2t = - ...
lambda^4 * (1 / prod(4*pi^2*((VV.*VV + 2*VV.*BB)))^0.5) * ...
exp(-0.5 * (pdist2(xx2sqFin,-xx2sqFin) + dist4)) .* invKxx;
YY = lambda^4 * ...
(1 / prod(4*pi^2*((VV.*VV + 2*VV.*BB)))^0.5) * ...
exp(-0.5 * (pdist2(xx2sqFin,-xx2sqFin) + dist4)) .* ...
postProd + sig2t;
% Mean of the integral at iteration 't', before scaling back up:
mu(t) = (aa + 0.5*(sum(YY(:))+1*lambda.^2 * (1/(prod(2*pi*VV)^0.5))) );
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% Variance of the integral, before scaling back up:
%---------------- Tmp Vars to calculate first term -------------------%
GG_coeff = lambda^6 * 1/prod(16*pi^4*(VV.*VV + 2*VV.*BB).*VV.*BB.* ...
(((VV.*VV + 2*VV.*BB)./BB) + ...
((VV.*VV + 2*VV.*BB)./VV) + VV + BB))^0.5 * ...
prod(2*pi*(VV.*VV + 2*VV.*BB))^0.5;
tmpVar = ((VV.*VV + 2*VV.*BB).*(Yvar + Yvar.*(VV./BB + VV + BB)));
xx3sq = xxIter .* ...
repmat(sqrt((Yvar.*VV)./tmpVar),currNumSamples,1);
bb3ch = bb .* sqrt((Yvar.*VV)./tmpVar);
xx2sqGGlh = xx2sqFin + pdist2_squared_fast(xx3sq,bb3ch);
xx4sq = xxIter .* ...
repmat(sqrt((VV.*Yvar)./tmpVar),currNumSamples,1);
bb4ch = bb .* sqrt((VV.*Yvar)./tmpVar);
xx5sq = xxIter .* ...
repmat(sqrt((VV.*Yvar.^2)./(tmpVar.*BB)),currNumSamples,1);
bb5ch = bb .* sqrt((VV.*Yvar.^2)./(tmpVar.*BB));
xx2sqGGrh = pdist2_squared_fast(xx4sq,bb4ch) + ...
pdist2_squared_fast(xx5sq,bb5ch);
xxIterScaled3 = xxIter .* ...
repmat(sqrt(Yvar.*BB./tmpVar),currNumSamples,1);
dist4 = pdist2_squared_fast(xxIterScaled3,xxIterScaled3);
xxIterScaled4 = xxIter .* ...
repmat(sqrt(BB./(Yvar.*VV)),currNumSamples,1);
dist5 = pdist2_squared_fast(xxIterScaled4,xxIterScaled4);
GG = GG_coeff * postProd .* ...
exp(-0.5*(pdist2(xx2sqGGlh, -xx2sqGGrh) + dist4));
YY2 = lambda^4 * (1 / prod(4*pi^2*((VV.*VV + 2*VV.*BB)))^0.5) * ...
exp(-0.5 * (pdist2(xx2sqFin,-xx2sqFin) + dist5)) .* ...
repmat(ww',length(ww),1);
Var(t) = (sum(sum(GG)) - sum(YY2,2)'*(invKxx * sum(YY2,2)));
%---------------- Tmp Vars to calculate second term ------------------%
tmp_2_mainvar = (BB.*(VV.*VV + 2*VV.*BB) + (((VV.*VV)./BB) + 2*VV) +...
(((VV.*VV+2*VV.*BB).*(VV.*VV+2*VV.*BB))./(VV)) + ...
VV.*(VV.*VV + 2*VV.*BB));
tmp_2_coeff = lambda^8 * 1/prod(8*pi^3*(VV.*VV+2*VV.*BB))^0.5 * ...
prod((VV+2*BB).*(VV.*VV./BB+2*VV))^0.5 * ...
1/prod(tmp_2_mainvar)^0.5;
ScaledVar2_0 = ((VV)./(VV.*VV+2*VV.*BB));
xxIterScaledVar2_0 = xxIter .* repmat(sqrt(ScaledVar2_0),currNumSamples,1);
bbScaledVar2_0 = bb .* sqrt(ScaledVar2_0);
distVar2_0 = pdist2_squared_fast(xxIterScaledVar2_0, bbScaledVar2_0);
ScaledVar2_1 = ((VV.*VV + 2*VV.*BB)./(tmp_2_mainvar));
xxIterScaledVar2_1 = xxIter .* repmat(sqrt(ScaledVar2_1),currNumSamples,1);
bbScaledVar2_1 = bb .* sqrt(ScaledVar2_1);
distVar2_1 = pdist2_squared_fast(xxIterScaledVar2_1, bbScaledVar2_1);
ScaledVar2_2 = ((VV.*VV + 2*VV.*BB).*(VV.*VV + 2*VV.*BB))./(tmp_2_mainvar.*(VV.*BB));
xxIterScaledVar2_2 = xxIter .* repmat(sqrt(ScaledVar2_2),currNumSamples,1);
bbScaledVar2_2 = bb .* sqrt(ScaledVar2_2);
distVar2_2 = pdist2_squared_fast(xxIterScaledVar2_2, bbScaledVar2_2);
ScaledVar2_3 = ScaledVar2_1;
xxIterScaledVar2_3 = xxIter .* repmat(sqrt(ScaledVar2_3),currNumSamples,1);
bbScaledVar2_3 = bb .* sqrt(ScaledVar2_3);
distVar2_3 = pdist2_squared_fast(xxIterScaledVar2_3, bbScaledVar2_3);
ScaledVar2_4 = ((VV.*BB)./(tmp_2_mainvar));
xxIterScaledVar2_4 = xxIter .* repmat(sqrt(ScaledVar2_4),currNumSamples,1);
bbScaledVar2_4 = bb .* sqrt(ScaledVar2_4);
distVar2_4 = pdist2_squared_fast(xxIterScaledVar2_4, bbScaledVar2_4);
distVar2lh = distVar2_1 + distVar2_2;
distVar2rh = distVar2_0 + distVar2_3 + distVar2_4;
distVar2 = pdist2(distVar2lh,-distVar2rh);
%---------------- Tmp Vars to calculate third term -------------------%
tmp_3_mainvar = VV.*VV + 2*VV.*BB;
tmp_3_coeff = lambda^12 * 1/prod(4*pi^2*tmp_3_mainvar);
ScaledVar3_0 = (VV./((VV.*VV+2*VV.*BB)));
xxIterScaledVar3_0 = xxIter .* repmat(sqrt(ScaledVar3_0),currNumSamples,1);
bbScaledVar3_0 = bb .* sqrt(ScaledVar3_0);
distVar3_0 = pdist2_squared_fast(xxIterScaledVar3_0, bbScaledVar3_0);
ScaledVar3_1 = (VV./(VV.*VV+2*VV.*BB));
xxIterScaledVar3_1 = xxIter .* repmat(sqrt(ScaledVar3_1),currNumSamples,1);
bbScaledVar3_1 = bb .* sqrt(ScaledVar3_1);
distVar3_1 = pdist2_squared_fast(xxIterScaledVar3_1, bbScaledVar3_1);
ScaledVar3_2 = (BB./(VV.*VV+2*VV.*BB));
xxIterScaledVar3_2 = xxIter .* repmat(sqrt(ScaledVar3_2),currNumSamples,1);
distVar3_2 = pdist2_squared_fast(xxIterScaledVar3_2, xxIterScaledVar3_2);
ScaledVar3_3 = (BB./(VV.*VV+2*VV.*BB));
xxIterScaledVar3_3 = xxIter .* repmat(sqrt(ScaledVar3_3),currNumSamples,1);
distVar3_3 = pdist2_squared_fast(xxIterScaledVar3_3, xxIterScaledVar3_3);
ScaledVar3_4 = (VV./(VV.*VV+2*VV.*BB));
xxIterScaledVar3_4 = xxIter .* repmat(sqrt(ScaledVar3_4),currNumSamples,1);
bbScaledVar3_4 = bb .* sqrt(ScaledVar3_4);
distVar3_4 = pdist2_squared_fast(xxIterScaledVar3_4, bbScaledVar3_4);
ScaledVar3_5 = (VV./(VV.*VV+2*VV.*BB));
xxIterScaledVar3_5 = xxIter .* repmat(sqrt(ScaledVar3_5),currNumSamples,1);
bbScaledVar3_5 = bb .* sqrt(ScaledVar3_5);
distVar3_5 = pdist2_squared_fast(xxIterScaledVar3_5, bbScaledVar3_5);
%----------------- Combine terms to get total var --------------------%
tmp_1 = lambda.^4/prod(8*pi^2*VV.*(0.5*(VV+2*BB)+BB))^0.5;
tmp_2 = invKxx .* (tmp_2_coeff * exp(-0.5*distVar2));
tmp_3 = tmp_3_coeff * exp(-0.5*distVar3_0)' * ...
(invKxx.*exp(-0.5*distVar3_2))* exp(-0.5*distVar3_1) * ...
exp(-0.5*distVar3_4)'*(invKxx.*exp(-0.5*distVar3_3)) * ...
exp(-0.5*distVar3_5);
Var(t) = Var(t) + 0.5*(tmp_1 - 2*sum(tmp_2(:)) + sum(tmp_3(:)));
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% Actively select next sample point:
if rand < 1.1 % Sample starting location for search from prior.
strtSamp = mvnrnd(bb,diag(BB),1);
else
strtSamp = 2*range(2,:).*rand(1,dim) - 50;
end
% If using local optimiser (fast):
%EV = @(x) expectedVarL( transp(x), lambda, VV, ...
% lHatD, xxIter, invKxx, jitterNoise, bb, BB ); %Utility function
%newX = fmincon( EV, strtSamp,[],[],[],[], ...
% range(1,:),range(2,:),[],options2 );
% If using global optimiser (cmaes):
newX = cmaes( 'expectedVarM', strtSamp', [],opts, lambda, VV, ...
lHatD, xxIter, invKxx, jitterNoise, bb, BB);
newX = newX';
xx(numSamples-currNumSamples,:) = newX;
clktime(t) = cputime - tmpT;
currNumSamples = currNumSamples + 1;
end
fprintf('\n done.\n');
log_mu = log(mu) + logscaling;
log_Var = log(Var) + 2*logscaling;
end