def genbiasedmat52ojf(d, lb, ub, xls, sls):
#s normalised to 0 exact, 1
from ESutils import gen_dataset
nt = 30
[X, Y, S, D] = gen_dataset(nt, d + 1, lb + [0], ub + [1], GPdc.MAT52,
sp.array([1.5] + [xls] * d + [sls]))
G = GPdc.GPcore(
X, Y, S, D,
GPdc.kernel(GPdc.MAT52, d + 1, sp.array([1.5] + [0.30] * d + [sls])))
def ojf(x, **ev):
#print "\nojfinput: {} : {}".format(x,ev)
dx = ev['d']
s = ev['s']
if ev['s'] > 0:
noise = sp.random.normal(scale=sp.sqrt(ev['s']))
else:
noise = 0
#print 'noise in ojf {}'.format(noise)
xa = ev['xa']
x = sp.array(x)
xfull = sp.hstack([x, xa])
return G.infer_m(xfull, [dx])[0, 0] + noise, 1., dict()
def dirwrap(x, y):
z = G.infer_m(sp.hstack(sp.array(x) + [0.]), [[sp.NaN]])[0, 0]
return (z, 0)
[xmin, ymin, ierror] = DIRECT.solve(dirwrap,
lb,
ub,
user_data=[],
algmethod=1,
maxf=40000,
logfilename='/dev/null')
#print [xmin, ymin]
def spowrap(x):
z = G.infer_m(sp.hstack(sp.array(x) + [0.]), [[sp.NaN]])[0, 0]
return z
y = spm(spowrap,
xmin,
method='l-bfgs-b',
bounds=[(-1, 1)] * d,
options={'ftol': 1e-15})
xmin = y.x
ymin = spowrap(y.x)
#print [xmin,ymin]
if sp.isnan(ymin):
logger.warning('generator got nan optimizing objective. retrying...')
return genbiasedmat52ojf(d, lb, ub, xls, sls)
logger.info(
'generated function xmin {} ymin {} globopt:{} locopt:{}'.format(
xmin, ymin, ierror, y.status))
return ojf, xmin, ymin
def gendecayingpositiveojf(d, lb, ub, sim):
# s normalised to 0 exact, 1
from ESutils import gen_dataset
nt = 20
cl = 2.
[X, Y, S, D] = gen_dataset(nt, d, lb, ub, GPdc.MAT52,
sp.array([1] + [0.30] * d))
G0 = GPdc.GPcore(X, Y, S, D,
GPdc.kernel(GPdc.MAT52, d, sp.array([1] + [0.30] * d)))
[X, Y, S, D] = gen_dataset(nt, d, lb, ub, GPdc.MAT52,
sp.array([1] + [0.30] * d))
G1 = GPdc.GPcore(X, Y, S, D,
GPdc.kernel(GPdc.MAT52, d, sp.array([1] + [0.30] * d)))
[X, Y, S, D] = gen_dataset(nt, d, lb, ub, GPdc.MAT52,
sp.array([1] + [0.30] * d))
G2 = GPdc.GPcore(X, Y, S, D,
GPdc.kernel(GPdc.MAT52, d, sp.array([1] + [0.30] * d)))
def p0(x):
v = G0.infer_m(x, [[sp.NaN]])[0, 0]
y = v + 1 if v > 0 else sp.exp(v)
return y
def p1(x):
v = G1.infer_m(x, [[sp.NaN]])[0, 0]
y = v + 1 if v > 0 else sp.exp(v)
return y
def p2(x):
v = G2.infer_m(x, [[sp.NaN]])[0, 0]
y = v + 1 if v > 0 else sp.exp(v)
return y
def ojf(x, **ev):
#print "ex: {} {}".format(x,ev['xa'])
# print "\nojfinput: {} : {}".format(x,ev)
dx = ev['d']
s = ev['s']
if ev['s'] > 0:
noise = sp.random.normal(scale=sp.sqrt(ev['s']))
else:
noise = 0
# print 'noise in ojf {}'.format(noise)
xa = ev['xa']
x = sp.array(x)
y0 = p0(x)
y1 = sim * p1(x) + y0
l = p2(x)
A = (y1 - y0) / (sp.exp(l) - 1)
B = y0 - A
y = A * sp.exp(l * xa) + B
c = sp.exp(-cl * xa)
return y + noise, c, dict()
def dirwrap(x, y):
z = ojf(x, **{'d': [sp.NaN], 's': 0, 'xa': 0})[0]
return (z, 0)
[xmin, ymin, ierror] = DIRECT.solve(dirwrap,
lb,
ub,
user_data=[],
algmethod=1,
maxf=20000,
logfilename='/dev/null')
# print [xmin, ymin]
def spowrap(x):
z = ojf(x, **{'d': [sp.NaN], 's': 0, 'xa': 0})[0]
return z
y = spm(spowrap, xmin, method='nelder-mead', options={'fatol': 1e-12})
xmin = y.x
ymin = spowrap(y.x)
# print [xmin,ymin]
logger.info(
'generated function xmin {} ymin {} yminisnan{} globopt:{} locopt:{}'.
format(xmin, ymin, sp.isnan(ymin), ierror, y.status))
if sp.isnan(ymin):
logger.warning('generator got nan optimizing objective. retrying...')
return gendecayingpositiveojf(d, lb, ub)
return ojf, xmin, ymin
def genmat52ojf(d, lb, ub, A=1., ls=0.3, fixs=-1, ki=GPdc.MAT52):
from ESutils import gen_dataset
if isinstance(ls, float):
ls = [ls] * d
#nt=sp.maximum(250,sp.minimum(int(d*100./sp.product(ls)),3000))
nt = sp.maximum(150, sp.minimum(int(d * 20. / sp.product(ls)), 2500))
#print(nt)
#nt=500
s = 1e-9
for s in sp.logspace(-9, 0, 20):
try:
[X, Y, S, D] = gen_dataset(nt,
d,
lb,
ub,
ki,
sp.array([A] + ls),
s=1e-9)
break
except:
raise
# from matplotlib import pyplot as plt
# plt.figure()
# plt.plot(sp.array(X),sp.array(Y),'b.')
# plt.show(block=True)
print('training GP')
G = GPdc.GPcore(X, Y, S, D, GPdc.kernel(ki, d, sp.array([A] + ls)))
def wrap(x):
xq = sp.copy(x)
xq.resize([1, d])
a = G.infer_m_post(xq, [[sp.NaN]])
return a[0, 0]
dpara = {
'user_data': [],
'algmethod': 1,
'maxf': 40000,
'logfilename': '/dev/null'
}
lpara = {'ftol': 1e-20, 'maxfun': 1200}
xmin, ymin, ierror = gpbo.core.optutils.twopartopt(wrap, lb, ub, dpara,
lpara)
print('init {} {}'.format(xmin, ymin))
for i in xrange(250):
p = sp.random.normal(size=d) * 1e-2
res = spm(wrap,
xmin + p,
method='L-BFGS-B',
bounds=tuple([(lb[j], ub[j]) for j in xrange(d)]),
options={'gtol': 1e-30})
# print(res)
#print(xmin,res.x,wrap(xmin),wrap(res.x)<wrap(xmin))
if wrap(res.x) < wrap(xmin):
xmin = res.x
# ymin = ymin+res.fun
print('change: {} {}'.format(xmin, wrap(res.x) - ymin))
ymin = wrap(xmin)
def ojf(x, **ev):
dx = ev['d']
s = ev['s']
if fixs < 0:
if ev['s'] > 0 and not 'cheattrue' in list(ev.keys()):
noise = sp.random.normal(scale=sp.sqrt(ev['s']))
else:
noise = 0
else:
if not 'cheattrue' in list(ev.keys()):
noise = sp.random.normal(scale=sp.sqrt(fixs))
else:
noise = 0.
y = wrap(x) + noise #G.infer_m(sp.array(x),[dx])[0,0]+noise
if not 'silent' in list(ev.keys()):
print('ojf at {} {} returned {} noise {}'.format([i for i in x],
ev, y, noise))
return y - ymin, 1., dict()
logger.info(
'generated function xmin {} ymin {}(shifted to 0.) opt:{}'.format(
xmin, ymin, ierror))
return ojf, xmin, 0.