-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathDMEA_method.py
327 lines (274 loc) · 13.3 KB
/
DMEA_method.py
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
'''
Code adapted from https://github.com/Alaya-in-Matrix/pyMACE
{pmlr-v80-lyu18a,
title = {Batch Bayesian Optimization via Multi-objective Acquisition Ensemble for Automated Analog Circuit Design},
author = {Lyu, Wenlong and Yang, Fan and Yan, Changhao and Zhou, Dian and Zeng, Xuan},
booktitle = {Proceedings of the 35th International Conference on Machine Learning},
pages = {3306--3314},
year = {2018},
editor = {Dy, Jennifer and Krause, Andreas},
volume = {80},
series = {Proceedings of Machine Learning Research},
month = {10--15 Jul},
publisher = {PMLR},
pdf = {http://proceedings.mlr.press/v80/lyu18a/lyu18a.pdf},
url = {https://proceedings.mlr.press/v80/lyu18a.html},
}
'''
from GP_model import GP
import numpy as np
from platypus import NSGAII, Problem, Real, Solution, InjectedPopulation, Archive
from scipy.optimize import fmin_l_bfgs_b
from smt.sampling_methods import LHS
import pandas as pd
import GPy
from prefer_select import Preferred_select
# eta is a hyperparameters
# lsx : the collected x in last round
# lsy : the test function value corresponding to the collected x in last round
class DMEA:
def __init__(self, f, lb, ub, num_init, max_iter, k, path, mo_eval=7000, func_name=None, eta=0):
"""
f: the objective function:
input: D row vector
output: scalar value
lb: lower bound
ub: upper bound
num_init: number of initial random sampling
max_iter: number of iterations
B: batch size, the total number of function evaluations would be num_init + B * max_iter
"""
self.f = f
self.lb = lb.reshape(lb.size)
self.ub = ub.reshape(ub.size)
self.dim = self.lb.size
self.num_init = num_init
self.max_iter = max_iter
self.k = k # batch size
self.mo_eval = mo_eval
self.func_name = func_name
self.eta = eta
self.path = path
self.ddbbxx = []
domain = []
for i in range(lb.size):
gg = (self.lb[i], self.ub[i])
domain.append({'name': f'x_{i + 1}', 'type': 'continuous', 'domain': gg})
self.domain = domain
def init(self):
# For initialization, the best self. B historical sampling points in the initial sampling are singled out
# as the historical sampling point of the recommended sampling points in previous round,
# so that the penalty information will be initialized at the first recommended point
# The recommended sample point information for the previous round is stored in self.lsx, self.lsy
# self.dbx stores all the X previously selected
# self.dby stores all the Y previously selected
self.dbx = np.zeros((self.num_init, self.dim))
# Latin hypercube sampling
xlimits = np.array([(self.lb[i], self.ub[i]) for i in range(self.dim)])
samples = LHS(xlimits=xlimits, random_state=1)
self.dbx = samples(self.num_init)
self.ddbbxx = samples(self.num_init)
self.dby = np.zeros((self.num_init, 1))
self.best_y = np.inf
self.min_y = []
self.min_index = []
for i in range(self.num_init):
y = self.f(self.dbx[i])
if y < self.best_y:
self.best_y = y
self.best_x = self.dbx[i]
self.dby[i] = y
# the best self. B points are taken out as the initial lsx, lxy
dbx = self.dbx.tolist()
dby = self.dby.tolist()
db = pd.DataFrame({'dbx': dbx, 'dby': dby})
self.lsy = np.array(db['dby'][db['dby'].rank(method='first') - self.k <= 1e-3].tolist()).reshape(-1, 1)
self.lsx = np.array(db['dbx'][db['dby'].rank(method='first') - self.k <= 1e-3].tolist())
self.exx = np.array(db.drop(db.index[[db['dby'].rank(method='first') - self.k <= 1e-3]])['dbx'].tolist())
self.exy = np.array(db.drop(db.index[[db['dby'].rank(method='first') - self.k <= 1e-3]])['dby'].tolist())
# Initialize the Gaussian model
mean_ = np.mean(self.exy)
std_ = np.std(self.exy)
kern = GPy.kern.Matern52(input_dim=self.dim, ARD=True)
self.m = GPy.models.GPRegression(self.exx, (self.exy - mean_) / std_, kern, noise_var=0)
def gen_guess(self):
num_guess = 1 + len(self.model.ms)
guess_x = np.zeros((num_guess, self.dim))
guess_x[0, :] = self.best_x
def obj(x, m):
m, _ = m.predict(x[None, :])
return m
def gobj(x, m):
dmdx, _ = m.predictive_gradients(x[None, :])
return dmdx
bounds = [(self.lb[i], self.ub[i]) for i in range(self.dim)]
for i in range(1, num_guess):
m = self.model.ms[i - 1]
xx = self.best_x + np.random.randn(self.best_x.size).reshape(self.best_x.shape) * 1e-3
print(np.random.randn(self.best_x.size))
def mobj(x):
return obj(x, m)
def gmobj(x):
return gobj(x, m)
x, _, _ = fmin_l_bfgs_b(mobj, xx, gmobj, bounds=bounds)
guess_x[i, :] = np.array(x)
return guess_x
def optimize(self):
self.best_y = np.min(self.dby)
self.P = np.zeros((1, 7)).ravel()
for iter in range(self.max_iter):
print('\n', self.func_name, f'{iter}/{self.max_iter}')
self.model = GP(iter=iter, train_x=self.dbx, train_y=self.dby, exx=self.exx, exy=self.exy, k=self.k,
lsx=self.lsx, lsy=self.lsy, P=self.P, f=self.f, domain=self.domain,
num_init=self.num_init, model=self.m, eta=self.eta)
self.P = self.model.CP
if iter == 0:
self.P = np.zeros((1, 7)).ravel()
self.exx = self.dbx
self.exy = self.dby
self.m = self.model.m
guess_x = self.gen_guess()
print('Hello')
num_guess = guess_x.shape[0]
self.log = []
# Build a multi-objective optimization problem
def obj(x):
val = []
obj_list = self.model.MACE_acq(np.array([x]))
for i in range(len(obj_list)):
self.log.append(obj_list[i][1])
if obj_list[i][1] == 1:
obj_list[i][0] = -1 * np.log(1e-40 + obj_list[i][0])
val.append(obj_list[i][0][0])
return val
# self.dim is the number of decision variables，
# 3 is the number of multi-objective optimization problem，That is, the number of acquisition functions：
# The acquisition functions are EI, LCB, and PI
problem = Problem(self.dim, 3)
for i in range(self.dim):
problem.types[i] = Real(self.lb[i], self.ub[i])
init_s = [Solution(problem) for i in range(num_guess)]
for i in range(num_guess):
init_s[i].variables = [x for x in guess_x[i, :]]
problem.function = obj
gen = InjectedPopulation(init_s)
arch = Archive()
# to Use the NSGAII algorithm to calculate multi-objective optimization
algorithm = NSGAII(problem, population=100, generator=gen, archive=arch)
algorithm.run(self.mo_eval)
if len(algorithm.result) > self.k:
optimized = algorithm.result
else:
optimized = algorithm.population
x_one_optimal = np.array(optimized[0].objectives)
dimention = x_one_optimal.shape[0]
trust_vector = self.model.T
assert len(trust_vector) == dimention
# X_pf Stored the pareto front of multi-objective acquisition functions optimized in this round
# X_ps Stored the pareto optimal corresponding to X_pf in the round
X_pf = np.array([np.array(optimized[i].objectives) for i in range(len(optimized))])
X_ps = np.array([np.array(optimized[i].variables) for i in range(len(optimized))])
# Use pandas to combine all X_pf and X_ps, and then delete the entire row
# (including the column of X_pf and X_ps) depending on whether the X_ps is duplicated or not
X_pf_ps = pd.DataFrame()
for i in range(X_pf.shape[1]):
X_pf_ps.loc[:, f'pf{i}'] = X_pf[..., i].flatten()
for i in range(X_ps.shape[1]):
X_pf_ps.loc[:, f'ps{i}'] = X_ps[..., i].flatten()
subset_ps = [f'ps{i}' for i in range(X_ps.shape[1])]
X_pf_ps.dropna(axis=0, how='any', inplace=True)
X_pf_ps = X_pf_ps.drop_duplicates(subset=subset_ps, keep="first")
# Remove duplicate elements from the X_pf
X_pf = X_pf_ps.iloc[:, 0:X_pf.shape[1]]
X_ps = X_pf_ps.iloc[:, -X_ps.shape[1]:]
X_pf = np.array(X_pf)
X_ps = np.array(X_ps)
X_c = []
if len(X_ps) > self.k:
repeat_num = 0
# Merge the three pareto optimals selected in this round and self.dbx,
# and then check whether repeat or not and delete.
# comparing the newly merged self.dbx before and after the deletion,if the length of the array changes,
# the difference before and after array length is noted as repeat_num and assigned to
# the Preferred select strategy
self_dbx_pd = pd.DataFrame()
for i in range(self.dbx.shape[1]):
self_dbx_pd.loc[:, f'dbx{i}'] = self.dbx[..., i].flatten()
subset_dbx = [f'dbx{i}' for i in range(self.dbx.shape[1])]
self_dbx_pd = self_dbx_pd.drop_duplicates(subset=subset_dbx, keep="first")
self.dbx = self_dbx_pd.to_numpy()
for ii in range(dimention):
a = X_pf[:, ii]
list_a = a.tolist()
min_index = list_a.index(min(list_a))
x = np.array(X_ps[min_index])
self.dbx = np.append(self.dbx, x.reshape(1, x.size), axis=0)
len_dbx_bef = len(self.dbx)
dbx_pd = pd.DataFrame()
for i in range(self.dbx.shape[1]):
dbx_pd.loc[:, f'dbx{i}'] = self.dbx[..., i].flatten()
subset_dbx = [f'dbx{i}' for i in range(self.dbx.shape[1])]
dbx_pd = dbx_pd.drop_duplicates(subset=subset_dbx, keep="first")
self.dbx = dbx_pd.to_numpy()
len_dbx_after = len(self.dbx)
if len_dbx_after != len_dbx_bef:
repeat_num += 1
else:
X_c.append(min_index)
# Preferred select strategy
X_reco_pf = Preferred_select(self.k - dimention + repeat_num, trust_vector, X_pf, X_ps, self.dbx)
for k in X_reco_pf:
X_c.append(k)
prefer_x = X_reco_pf
for i in prefer_x:
x = np.array(X_ps[i])
self.dbx = np.append(self.dbx, x.reshape(1, x.size), axis=0)
elif len(X_ps) == self.k:
for i in range(len(X_ps)):
X_c.append(i)
for i in X_c:
x = np.array(X_ps[i])
self.dbx = np.append(self.dbx, x.reshape(1, x.size), axis=0)
else:
for i in range(len(X_ps)):
X_c.append(i)
for i in range(self.k - len(X_ps)):
X_c.append(0)
for i in X_c:
x = np.array(X_ps[i])
self.dbx = np.append(self.dbx, x.reshape(1, x.size), axis=0)
lsx = []
lsy = []
# evaluate the selected B points
for i in X_c:
x = np.array(X_ps[i])
y = self.f(x) # evaluation
lsx.append(x)
lsy.append(y)
if y < self.best_y:
self.best_y = y
self.best_x = x
self.ddbbxx = np.append(self.ddbbxx, x.reshape(1, x.size), axis=0)
self.dby = np.append(self.dby, y.reshape(1, 1), axis=0)
self.dbx = self.ddbbxx
self.lsx = np.array(lsx)
self.lsy = np.array(lsy).reshape(-1, 1)
# Save the Pareto front and solution in each iteration
pf = np.array([s.objectives for s in optimized])
ps = np.array([s.variables for s in optimized])
self.pf = pf
self.ps = ps
np.savetxt(
f'{self.path[1]}/dbx_Batchsize_{self.k}_{self.func_name}__eta_{self.eta}.txt',
self.dbx)
np.savetxt(
f'{self.path[1]}/dby_Batchsize_{self.k}_{self.func_name}__eta_{self.eta}.txt',
self.dby)
# Save the smallest test function value in each iteration
min_y_2 = self.dby.min()
self.min_y.append(min_y_2)
min_x_index = (np.where(self.dby == min_y_2)[0][0])
min_index_2 = np.concatenate((np.array([min_y_2]), self.dbx[min_x_index]), axis=0)
self.min_index.append(min_index_2)
np.savetxt(f'{self.path[0]}/min_y_' + f'Batchsize{self.k}' + f'max_iter{self.max_iter}' + f'_{self.func_name}' + f'_eta_{self.eta}' + '.txt',
self.min_index, delimiter=',', fmt='%s')