def test_bisect():
xtol = 1e-6
## simple tests
root = noisyopt.bisect(lambda x: x, -2, 2, xtol=xtol)
npt.assert_allclose(root, 0.0, atol=xtol)
root = noisyopt.bisect(lambda x: x-1, -2, 2, xtol=xtol)
npt.assert_allclose(root, 1.0, atol=xtol)
## extrapolate if 0 outside of interval
root = noisyopt.bisect(lambda x: x, 1, 2, xtol=xtol)
npt.assert_allclose(root, 0.0, atol=xtol)
npt.assert_raises(noisyopt.BisectException,
noisyopt.bisect, lambda x: x, 1, 2,
xtol=xtol, outside='raise')
## extrapolate with nonlinear function
root = noisyopt.bisect(lambda x: x+x**2, 1.0, 2, xtol=xtol)
assert root < 1.0
## test with stochastic function
xtol = 1e-1
func = lambda x: x - 0.25 + np.random.normal(scale=0.01)
root = noisyopt.bisect(noisyopt.AveragedFunction(func), -2, 2, xtol=xtol,
errorcontrol=True)
npt.assert_allclose(root, 0.25, atol=xtol)
def get_mde_size_sample(self):
n_null_metric_samples = 2000
n_alt_metric_samples = 400
null_metric_samples = []
# Metric distribution under null hypothesis
for sample in range(0, n_null_metric_samples):
group_a1_responses = self.get_group_a1_samples()
group_a2_responses = self.get_group_a2_samples()
group_b1_responses = self.get_group_b1_samples()
group_b2_responses_null = self.get_group_b2_samples_by_effect(effect=0)
null_metric_samples.append(
(np.mean(group_b2_responses_null) - np.mean(group_b1_responses)) -
(np.mean(group_a2_responses) - np.mean(group_a1_responses))
)
null_critical_value = np.percentile(null_metric_samples, (1 - self.alpha / 2) * 100)
# Speeding up the search by bounding the search space more efficiently
indicative_mde = (
null_critical_value / norm.ppf(1 - (self.alpha / 2)) *
(norm.ppf(1 - (self.alpha / 2)) - norm.ppf(1 - self.pi_min)))
search_lbound = indicative_mde * 0.7
search_ubound = indicative_mde * 1.5
# Let the root finding tolerance to be around 0.05% of the difference between
# the group A means (in absolute value)
# With enough bootstrap samples this error would be further reduced
mde_search_tol = np.abs(np.mean(self.get_group_a2_samples()) - np.mean(self.get_group_a1_samples())) * 0.0005
# Extrapolation sometimes give extremely large (both +ve/-ve) values
# due to small denominator, the while loop guards against that
mde_sample = 2 * search_ubound
while np.abs(mde_sample) > search_ubound:
mde_sample = noisyopt.bisect(
noisyopt.AveragedFunction(
lambda x: self._simulated_power_dual_control(x, null_critical_value, n_alt_metric_samples) -
self.pi_min,
N=20),
search_lbound, search_ubound,
xtol=mde_search_tol, errorcontrol=True,
testkwargs={'alpha': 0.01, 'force': True, 'eps': 0.001, 'maxN': 160},
outside='extrapolate', ascending=True, disp=False)
return mde_sample
def test_bisect():
xtol = 1e-6
## simple tests
# ascending
root = noisyopt.bisect(lambda x: x, -2, 2, xtol=xtol, errorcontrol=False)
npt.assert_allclose(root, 0.0, atol=xtol)
root = noisyopt.bisect(lambda x: x - 1,
-2,
2,
xtol=xtol,
errorcontrol=False)
npt.assert_allclose(root, 1.0, atol=xtol)
# descending
root = noisyopt.bisect(lambda x: -x, -2, 2, xtol=xtol, errorcontrol=False)
npt.assert_allclose(root, 0.0, atol=xtol)
## extrapolate if 0 outside of interval
root = noisyopt.bisect(lambda x: x, 1, 2, xtol=xtol, errorcontrol=False)
npt.assert_allclose(root, 0.0, atol=xtol)
npt.assert_raises(noisyopt.BisectException,
noisyopt.bisect,
lambda x: x,
1,
2,
xtol=xtol,
outside='raise',
errorcontrol=False)
## extrapolate with nonlinear function
root = noisyopt.bisect(lambda x: x + x**2,
1.0,
2,
xtol=xtol,
errorcontrol=False)
assert root < 1.0
## test with stochastic function
xtol = 1e-1
func = lambda x: x - 0.25 + np.random.normal(scale=0.01)
root = noisyopt.bisect(noisyopt.AveragedFunction(func),
-2,
2,
xtol=xtol,
errorcontrol=True)
npt.assert_allclose(root, 0.25, atol=xtol)
sig_fits.append(guess)
logMmin_list.append(m)
ng_diff = wp_ng_vals[0] - ng
func_val.append(ng_diff)
return ng_diff
for m in logMmin:
print("logMmin: " + str(np.log10(m)))
model_instance.param_dict['logMmin'] = np.log10(m)
#res = minimizeCompass(_find_sigma,[0.3], bounds = [[0.01,1.5]],deltatol=0.0001,
# niter = 100, paired=False, disp=False, errorcontrol=False)
#res = minimizeCompass(_find_sigma,[0.3], bounds = [[0.01,1.5]],
# niter = 100, paired=False, disp=False, errorcontrol=False)
avfunc = AveragedFunction(_find_sigma)
res = bisect(avfunc,
0.01,
2.,
xtol=1e-06,
errorcontrol=True,
outside='extrapolate',
ascending=None,
disp=False,
testkwargs={0.0001})
#res = op.minimize(_get_ng,[12.,0.30],options={'maxiter': 1e3,'disp':True},tol=1e-3)
#res = minimizeCompass(_get_ng,[12.36,0.38],bounds = [[11.0, 14.0], [0.01, 2.0]], deltatol=1e#-6, a=5.0, disp = True, paired=False)
np.save("logMmin_sigma_errctrl.npy",
np.array([logMmin_list, sig_fits, func_val]))
njob = int(sys.argv[1])
data = []
for i in progressbar(range(nbatch)):
n = (njob - 1) * nbatch + i
boundary, aenv = paramscomb[n]
if disp:
print 'boundary %s, aenv %s' % (boundary, aenv)
bisect_kwargs = dict(xtol=xtol,
errorcontrol=True,
testkwargs=dict(alpha=alpha),
ascending=False,
disp=disp)
fargs = lambda_, mus, cup, aenv, niter, nburnin
pienvbnd = noisyopt.bisect(
noisyopt.DifferenceFunction(cuts[boundary[0]],
cuts[boundary[1]],
fargs1=fargs,
fargs2=fargs,
paired=True), 0, 1, **bisect_kwargs)
data.append([
boundary, lambda_, mus, cup, aenv, niter, nburnin, deltainit,
deltatol,
np.log10(feps), boundtol, qboundtol, xtol, xtolbound, pienvbnd
])
columns = [
'boundary', 'lambda_', 'mus', 'cup', 'aenv', 'niter', 'nburnin',
'deltainit', 'deltatol', 'logfeps', 'boundtol', 'qboundtol', 'xtol',
'xtolbound', 'pienvbnd'
]
np.savez_compressed(datadir + 'scan_phases_%g' % (njob),
data=data,
columns=columns)
# This is a minimal usage example for the root finding algorithm
import numpy as np
from noisyopt import bisect, AveragedFunction
# definition of noisy function of which root should be found
def func(x):
return x + 0.1*np.random.randn()
avfunc = AveragedFunction(func)
root = bisect(avfunc, -2, 2)
print(root)
('io', filterarray(aenvs, 0.01, 1.0)),
('pm', aenvs),
('pi', filterarray(aenvs, 0.0, 0.9)),
('po', aenvs[[0, -1]]),
('pc', aenvs),
])
# define cut functions
cuts = strategies.cuts(coptkwargs=coptkwargs, moptkwargs=moptkwargs,
ioptkwargs=ioptkwargs, poptkwargs=poptkwargs)
# run simulations
if parametercheck(datadir, sys.argv, paramscomb, nbatch):
njob = int(sys.argv[1])
data = []
for i in progressbar(range(nbatch)):
n = (njob-1) * nbatch + i
boundary, aenv = paramscomb[n]
if disp:
print 'boundary %s, aenv %s' % (boundary, aenv)
bisect_kwargs = dict(xtol=xtol, errorcontrol=True, testkwargs=dict(alpha=alpha),
ascending=False, disp=disp)
fargs = lambda_, mus, cup, aenv, niter, nburnin
pienvbnd = noisyopt.bisect(noisyopt.DifferenceFunction(cuts[boundary[0]], cuts[boundary[1]],
fargs1=fargs, fargs2=fargs, paired=True),
0, 1, **bisect_kwargs)
data.append([boundary, lambda_, mus, cup, aenv, niter, nburnin, deltainit, deltatol,
np.log10(feps), boundtol, qboundtol, xtol, xtolbound, pienvbnd])
columns = ['boundary', 'lambda_', 'mus', 'cup', 'aenv', 'niter', 'nburnin', 'deltainit',
'deltatol', 'logfeps', 'boundtol', 'qboundtol', 'xtol', 'xtolbound', 'pienvbnd']
np.savez_compressed(datadir + 'scan_phases_%g' % (njob), data=data, columns=columns)