def trunc_norm_prob(center):
""" get probability mass """
return (truncnorm.cdf(center + radius,
a=(low - mean) / radius,
b=(high - mean) / radius,
loc=mean,
scale=radius) -
truncnorm.cdf(center - radius,
a=(low - mean) / radius,
b=(high - mean) / radius,
loc=mean,
scale=radius))
def get_mfd(self, slip, area, shear_modulus=30.0):
'''
Calculates activity rate on the fault
:param float slip:
Slip rate in mm/yr
:param fault_width:
Width of the fault (km)
:param float disp_length_ratio:
Displacement to length ratio (dimensionless)
:param float shear_modulus:
Shear modulus of the fault (GPa)
:returns:
* Minimum Magnitude (float)
* Bin width (float)
* Occurrence Rates (numpy.ndarray)
'''
# Working in Nm so convert: shear_modulus - GPa -> Nm
# area - km ** 2. -> m ** 2.
# slip - mm/yr -> m/yr
moment_rate = (shear_modulus * 1.E9) * (area * 1.E6) * (slip / 1000.)
moment_mag = _scale_moment(self.mmax, in_nm=True)
characteristic_rate = moment_rate / moment_mag
if self.sigma and (fabs(self.sigma) > 1E-5):
self.mmin = self.mmax + (self.lower_bound * self.sigma)
mag_upper = self.mmax + (self.upper_bound * self.sigma)
mag_range = np.arange(self.mmin, mag_upper + self.bin_width,
self.bin_width)
self.occurrence_rate = characteristic_rate * (truncnorm.cdf(
mag_range + (self.bin_width / 2.),
self.lower_bound,
self.upper_bound,
loc=self.mmax,
scale=self.sigma) - truncnorm.cdf(mag_range -
(self.bin_width / 2.),
self.lower_bound,
self.upper_bound,
loc=self.mmax,
scale=self.sigma))
else:
# Returns only a single rate
self.mmin = self.mmax
self.occurrence_rate = np.array([characteristic_rate], dtype=float)
return self.mmin, self.bin_width, self.occurrence_rate
def fit(self,
model,
num_samples,
data,
truncated_lower=0.0,
truncated_upper=2.0,
threshold=0.01,
**kwargs):
start = time.time()
samples = self.sampler.sample(model, ut.random_normal, num_samples,
data, ut.mae_loss, **kwargs)
"""
The standard form of this distribution is a standard normal truncated
to the range [a, b] — notice that a and b are defined over the domain
of the standard normal. To convert clip values for a specific mean and
standard deviation, use:
a, b = (myclip_a - my_mean) / my_std, (myclip_b - my_mean) / my_std
"""
mean = np.mean(samples)
std = np.std(samples)
a = (truncated_lower - mean) / std
b = (truncated_upper - mean) / std
p_threshold = truncnorm.cdf(threshold, a, b, loc=mean, scale=std)
log_p = np.log(p_threshold)
sampling_time = time.time() - start
return {
'p': p_threshold,
'log_p': log_p,
'mean': mean,
'std': std,
'sampling_time': sampling_time,
'samples': samples
}
def cdf(self, x) -> float:
"""
Calculate the Normal cumulative distribution value at position `x`.
:param x: value where the cumulative distribution function is evaluated.
:return: value of the cumulative distribution function.
"""
if self.hard_clip_min is not None and (x < self.hard_clip_min):
return 0.
if self.hard_clip_max is not None and (x > self.hard_clip_max):
return 1.
if self.hard_clip_min is not None or self.hard_clip_max is not None:
a = -np.inf
b = np.inf
if self.hard_clip_min is not None:
a = (self.hard_clip_min - self.mean) / self.std
if self.hard_clip_max is not None:
b = (self.hard_clip_max - self.mean) / self.std
return truncnorm.cdf(x, a=a, b=b, loc=self.mean, scale=self.std)
return norm.cdf(x, loc=self.mean, scale=self.std)
def relative_seg(self, roi):
lower, upper = 0.33, 0.60
mu, std = 0.42, 0.06
a, b = (lower - mu) / std, (upper - mu) / std
return truncnorm.cdf(roi / np.max(roi), a, b, loc=mu, scale=std)
def cdf(self, x):
cdfs = truncnorm.cdf(x,
self.a,
self.b,
loc=self.means,
scale=self.sigmas)
return np.sum(np.dot(cdfs, self.coeff))
def cdf(self, x: Tuple[float]):
"""Find the CDF for a certain x value.
Args:
x (float): The value for which the CDF is needed.
Returns:
float: The CDF value at point x.
"""
x_a, x_b = (self.x_lower_bound - self.x_mean) / self.x_std, (
self.x_upper_bound - self.x_mean) / self.x_std
y_a, y_b = (self.y_lower_bound - self.y_mean) / self.y_std, (
self.y_upper_bound - self.y_mean) / self.y_std
return truncnorm.cdf(x[0], x_a, x_b, self.x_mean,
self.x_std) * truncnorm.cdf(
x[1], y_a, y_b, self.y_mean, self.y_std)
def CDF_func(self, t):
return truncnorm.cdf(
t,
self.__low_bound,
self.__up_bound,
loc=self.mu,
scale=self.sigma)
def absolute_seg(self, roi):
lower, upper = 2.0, 4.0
mu, std = 2.5, 0.5
a, b = (lower - mu) / std, (upper - mu) / std
return truncnorm.cdf(roi, a, b, loc=mu, scale=std)
def get_mfd(self, slip, area, shear_modulus=30.0):
'''
Calculates activity rate on the fault
:param float slip:
Slip rate in mm/yr
:param fault_width:
Width of the fault (km)
:param float disp_length_ratio:
Displacement to length ratio (dimensionless)
:param float shear_modulus:
Shear modulus of the fault (GPa)
:returns:
* Minimum Magnitude (float)
* Bin width (float)
* Occurrence Rates (numpy.ndarray)
'''
# Working in Nm so convert: shear_modulus - GPa -> Nm
# area - km ** 2. -> m ** 2.
# slip - mm/yr -> m/yr
moment_rate = (shear_modulus * 1.E9) * (area * 1.E6) * (slip / 1000.)
moment_mag = _scale_moment(self.mmax, in_nm=True)
characteristic_rate = moment_rate / moment_mag
if self.sigma and (fabs(self.sigma) > 1E-5):
self.mmin = self.mmax + (self.lower_bound * self.sigma)
mag_upper = self.mmax + (self.upper_bound * self.sigma)
mag_range = np.arange(self.mmin,
mag_upper + self.bin_width,
self.bin_width)
self.occurrence_rate = characteristic_rate * (
truncnorm.cdf(mag_range + (self.bin_width / 2.),
self.lower_bound, self.upper_bound,
loc=self.mmax, scale=self.sigma) -
truncnorm.cdf(mag_range - (self.bin_width / 2.),
self.lower_bound, self.upper_bound,
loc=self.mmax, scale=self.sigma))
else:
# Returns only a single rate
self.mmin = self.mmax
self.occurrence_rate = np.array([characteristic_rate], dtype=float)
return self.mmin, self.bin_width, self.occurrence_rate
def pretty_print(self):
a,b = (0-self.constraint.mean)/self.constraint.std,(1e6-self.constraint.mean)/self.constraint.std
ub_survival = truncnorm.sf(self.allocated_ub, a,b,loc=self.constraint.mean, scale=self.constraint.std)
lb_mass = truncnorm.cdf(self.allocated_lb, a,b,loc=self.constraint.mean, scale=self.constraint.std)
print(self.constraint.name + ": [" + str(self.allocated_lb) + "," + str(
self.allocated_ub) + "] (Risk: " + str(lb_mass+ub_survival) + ")")
def trunc_visualization(parameters):
[mu, sig, min, max] = parameters
a, b, = (min - mu) / sig, (max - mu) / sig
x_range = np.linspace(0, 1, 1000)
fig, ax = plt.subplots()
sns.lineplot(x_range, truncnorm.pdf(x_range, a, b, loc=mu, scale=sig), label='pdf')
sns.lineplot(x_range, truncnorm.cdf(x_range, a, b, loc=mu, scale=sig), label='cdf')
ax.legend()
return ax
def cdf(self, x: float):
"""Find the CDF for a certain x value.
Args:
x (float): The value for which the CDF is needed.
"""
a, b = (self.lower_bound - self.mean) / self.std, (
self.upper_bound - self.mean) / self.std
return truncnorm.cdf(x, a, b, self.mean, self.std)
def absolute_seg(self, roi):
# lower, upper = 2.0, 4.0
# # mu, std = 2.5, 0.5
# mu, std = 3.0, 0.5
lower, upper = self.tvals_probs['absolute']['lower'], self.tvals_probs[
'absolute']['upper']
mu, std = self.tvals_probs['absolute']['mu'], self.tvals_probs[
'absolute']['std']
a, b = (lower - mu) / std, (upper - mu) / std
return truncnorm.cdf(roi, a, b, loc=mu, scale=std)
def get_mfd(self, slip, shear_modulus, area):
'''Calculates activity rate'''
self.mfd_params['mmin'] = self.mfd_params['mmax'] + (
self.mfd_params['Lower'] * self.mfd_params['Sigma'])
moment_rate = (shear_modulus * 1.E9) * (area * 1.E6) * (slip / 1000.)
mag_upper = self.mfd_params['mmax'] + (self.mfd_params['Upper'] *
self.mfd_params['Sigma'])
mag_range = np.arange(self.mfd_params['mmin'],
mag_upper + self.bin_width + 1E-7,
self.bin_width)
#moment_mag = 10. ** (1.5 * mag_range + 9.05)
moment_mag = 10. ** (1.5 * self.mfd_params['mmax'] + 9.05)
characteristic_rate = moment_rate / moment_mag
self.occurrence_rate = characteristic_rate * (truncnorm.cdf(
mag_range + (self.bin_width / 2.), self.mfd_params['Lower'],
self.mfd_params['Upper'], loc=self.mfd_params['mmax'],
scale=self.mfd_params['Sigma']) - truncnorm.cdf(
mag_range - (self.bin_width / 2.), self.mfd_params['Lower'],
self.mfd_params['Upper'], loc=self.mfd_params['mmax'],
scale=self.mfd_params['Sigma']))
def relative_seg(self, roi):
# lower, upper = 0.33, 0.60
# mu, std = 0.42, 0.06
lower, upper = self.tvals_probs['relative']['lower'], self.tvals_probs[
'relative']['upper']
mu, std = self.tvals_probs['relative']['mu'], self.tvals_probs[
'relative']['std']
a, b = (lower - mu) / std, (upper - mu) / std
return truncnorm.cdf(roi / np.max(roi), a, b, loc=mu, scale=std)
def get_ci(self, stat, eta, l_thres, u_thres, alpha):
"""Calculation of one two-sided confidence interval"""
sigma = np.dot(eta, np.dot(self.cov, eta))
scale = np.sqrt(sigma)
pivot = lambda mu: truncnorm.cdf(stat, (l_thres - mu) / scale,
(u_thres - mu) / scale,
loc=mu,
scale=scale)
lb = stat - 20. * scale # lower bound
ub = stat + 20. * scale # upper bound
ci_l = helper.find_root(pivot, 1 - alpha / 2, lb, ub)
ci_u = helper.find_root(pivot, alpha / 2, lb, ub)
return np.array([ci_l, ci_u])
def usrf(status, x, needF, neF, F, needG, neG, G, cu, iu, ru):
"""
==================================================================
Computes the nonlinear objective and constraint terms for the
problem.
==================================================================
"""
# print('called usrfun with ' + str(len(G)) + ' non-linear variables')
if (needF[0] != 0):
# the second last row is for chance constraint
F[neF[0] - 2] = 0
if cc_var > 0:
F[neF[0] - 2] += x[cc_var]
for idx in range(0, int(len(G) / 2)):
mean = prob_means[idx]
sigma = prob_stds[idx]
lb_var = prob_vars[2 * idx]
ub_var = prob_vars[2 * idx + 1]
# print("Mean: " + str(mean) + " / Sigma: " + str(sigma))
a, b = (0 - mean) / sigma, (1e6 - mean) / sigma
ub_survival = truncnorm.sf(x[ub_var], a, b, loc=mean, scale=sigma)
lb_mass = truncnorm.cdf(x[lb_var], a, b, loc=mean, scale=sigma)
F[neF[0] - 2] += ub_survival + lb_mass
# print('Updating F['+str(neF[0] - 2)+']: ' + str(x[lb_var]) + '-' + str(x[ub_var]) + ': ' + str(lb_mass) + "+" +str(ub_survival) + "="+str(F[neF[0] - 2]))
if (needG[0] != 0):
# Compute the partial derivatives of the chance constraint
# over the lower and upper bounds of the
# probabilistic durations
for idx in range(0, int(len(G) / 2)):
mean = prob_means[idx]
sigma = prob_stds[idx]
lb_var = prob_vars[2 * idx]
ub_var = prob_vars[2 * idx + 1]
a, b = (0 - mean) / sigma, (1e6 - mean) / sigma
# For the lower bound, the derivative is the Gaussian pdf
G[2 * idx] = truncnorm.pdf(x[lb_var], a, b, loc=mean, scale=sigma)
# For the upper bound, it is the negation of the Gaussian pdf
G[2 * idx +
1] = -1 * truncnorm.pdf(x[ub_var], a, b, loc=mean, scale=sigma)
def usrf(status, x, needF, neF, F, needG, neG, G, cu, iu, ru):
"""
==================================================================
Computes the nonlinear objective and constraint terms for the
problem.
==================================================================
"""
# print('called usrfun with ' + str(len(G)) + ' non-linear variables')
if (needF[0] != 0):
# the second last row is for chance constraint
F[neF[0] - 2] = 0
if cc_var > 0:
F[neF[0] - 2] += x[cc_var]
for idx in range(0, int(len(G)/ 2)):
mean = prob_means[idx]
sigma = prob_stds[idx]
lb_var = prob_vars[2 * idx]
ub_var = prob_vars[2 * idx+1]
# print("Mean: " + str(mean) + " / Sigma: " + str(sigma))
a, b = (0 - mean) / sigma, (1e6 - mean) / sigma
ub_survival = truncnorm.sf(x[ub_var],a,b, loc=mean, scale=sigma)
lb_mass = truncnorm.cdf(x[lb_var],a,b, loc=mean, scale=sigma)
F[neF[0] - 2] += ub_survival + lb_mass
# print('Updating F['+str(neF[0] - 2)+']: ' + str(x[lb_var]) + '-' + str(x[ub_var]) + ': ' + str(lb_mass) + "+" +str(ub_survival) + "="+str(F[neF[0] - 2]))
if (needG[0] != 0):
# Compute the partial derivatives of the chance constraint
# over the lower and upper bounds of the
# probabilistic durations
for idx in range(0, int(len(G) / 2)):
mean = prob_means[idx]
sigma = prob_stds[idx]
lb_var = prob_vars[2 * idx]
ub_var = prob_vars[2 * idx + 1]
a, b = (0 - mean) / sigma, (1e6 - mean) / sigma
# For the lower bound, the derivative is the Gaussian pdf
G[2 * idx] = truncnorm.pdf(x[lb_var], a,b, loc=mean, scale=sigma)
# For the upper bound, it is the negation of the Gaussian pdf
G[2 * idx + 1] = -1 * truncnorm.pdf(x[ub_var], a,b, loc=mean, scale=sigma)
def cdf(self,dat):
'''
Evaluates the cumulative distribution function on the data points in dat.
:param dat: Data points for which the c.d.f. will be computed.
:type dat: natter.DataModule.Data
:returns: A numpy array containing the probabilities.
:rtype: numpy.array
'''
#print dat.X
a,b = (self.param['a']-self.param['mu'])/self.param['sigma'],(self.param['b']-self.param['mu'])/self.param['sigma']
return squeeze(truncnorm.cdf(dat.X,a,b,loc=self.param['mu'],scale=self.param['sigma']))
def beliefs_loglike_binary(responses, signals, sm, wp, noise_type, noise):
belief_matrix = fwd.calc_belief_matrix(sm, wp)
probs = np.empty(len(signals))
for k in range(len(signals)):
bayes_s0 = belief_matrix[0][signals[k]]
given_s0 = responses[k]
#probs[k] = prob_noisy(given_s0, bayes_s0, noise_type, noise[k])
probs0 = truncnorm.cdf(.5, -bayes_s0 / noise, (1-bayes_s0)/scale, loc=bayes_s0,
scale=noise) #assuming truncnorm, write own func.
if responses[k] == 0:
probs[k] = probs0
else:
probs[k] = 1 - probs0
loglike = np.sum(np.log(probs))
return loglike
def MLE_X_trunc(self,low_support,high_support,threshold,x2nd,x_v,w_v):
'''
I do not know this MLE est
I just realized that I can genreate the conditional chain rule to calculate
the MLE (Because I calculate the probability!!!)
'''
self.setup_para(0)
low_support = low_support.reshape(self.N,1)
high_support = high_support.reshape(self.N,1)
low_support[-2]=low_support[-2]-0.05
high_support[-2]=high_support[-2]+0.05
x_flag1 = x_v >= low_support
x_flag2 = x_v <= high_support
check_flag_v1=np.prod(x_flag1, axis=0)
check_flag_v2=np.prod(x_flag2, axis=0)
check_flag_v1=check_flag_v1*check_flag_v2
nominator = np.sum(check_flag_v1*w_v)
denominator = np.sum(w_v)
mu=self.MU[-2]
sigma=self.SIGMA2[-2,-2]**0.5
density_2nd = truncnorm.pdf((x2nd-mu)/sigma,(threshold[-2]-mu)/sigma,10)
prob_1st = 1 - truncnorm.cdf((low_support[-1]-mu)/sigma,(threshold[-1]-mu)/sigma,10)
with np.errstate(divide='raise'):
try:
log_Prob = density_2nd + prob_1st+ np.log(nominator)-np.log(denominator)
except Exception as e:
print('-----------------------------------------------')
print('0 in log at {} bidders with {} reserve price'.format(self.N,threshold[0]))
print(low_support.flatten())
print(high_support.flatten())
print("density_2d: {0:.4}\t| prob_1st: {0:.4}\t| nominator: {0:.4}\t| denominator: {0:.4}\t| ".format(density_2nd,prob_1st,nominator,denominator))
log_Prob = np.nan
log_Prob = np.log(nominator)-np.log(denominator) + np.log(density_2nd) + np.log(prob_1st)
return log_Prob
def pretty_print(self):
a, b = (0 - self.constraint.mean) / self.constraint.std, (
1e6 - self.constraint.mean) / self.constraint.std
ub_survival = truncnorm.sf(self.allocated_ub,
a,
b,
loc=self.constraint.mean,
scale=self.constraint.std)
lb_mass = truncnorm.cdf(self.allocated_lb,
a,
b,
loc=self.constraint.mean,
scale=self.constraint.std)
print(self.constraint.name + ": [" + str(self.allocated_lb) + "," +
str(self.allocated_ub) + "] (Risk: " +
str(lb_mass + ub_survival) + ")")
def fit(self,
model,
num_samples,
data,
truncated_lower=0.0,
truncated_upper=2.0,
threshold=0.01,
**kwargs):
start = time.time()
samples, _ = self.sampler.sample(model, ut.random_normal, num_samples,
data, 'mae', **kwargs)
"""
The standard form of this distribution is a standard normal truncated
to the range [a, b] — notice that a and b are defined over the domain
of the standard normal. To convert clip values for a specific mean and
standard deviation, use:
a, b = (myclip_a - my_mean) / my_std, (myclip_b - my_mean) / my_std
"""
mean = np.mean(samples)
std = np.std(samples)
a = (truncated_lower - mean) / std
b = (truncated_upper - mean) / std
p_threshold = truncnorm.cdf(threshold, a, b, loc=mean, scale=std)
if p_threshold == 0:
# log_p = np.finfo(float).min
# Some software, e.g., mipego, have trouble dealing with long floats...
log_p = np.log(1e-300)
else:
log_p = np.log(p_threshold)
sampling_time = time.time() - start
return {
'p': p_threshold,
'log_p': log_p,
'mean': mean,
'std': std,
'sampling_time': sampling_time,
'samples': samples
}
def fit(self,
model,
num_samples,
x_df,
y_df,
truncated_lower=0.0,
truncated_upper=2.0,
threshold=0.01,
**kwargs):
samples = self.sampler.sample(model,
ut.random_normal,
num_samples,
x_df,
y_df,
ut.mae_loss,
**kwargs)
"""
The standard form of this distribution is a standard normal truncated
to the range [a, b] — notice that a and b are defined over the domain
of the standard normal. To convert clip values for a specific mean and
standard deviation, use:
a, b = (myclip_a - my_mean) / my_std, (myclip_b - my_mean) / my_std
"""
mean = np.mean(samples)
std = np.std(samples)
a = (truncated_lower - mean) / std
b = (truncated_upper - mean) / std
p_threshold = truncnorm.cdf(threshold,
a,
b,
loc=mean,
scale=std)
log_p = np.log(p_threshold)
return {'p': p_threshold,
'log_p': log_p,
'mean': mean,
'std': std,
'samples': samples}
def prob_distribution(self, task):
x_axis = np.arange(0, task.expiry_ + self.time_unit, self.time_unit)
a = task.deadline_
b = task.expiry_
probability = truncnorm.pdf(x_axis,
-b,
b,
loc=a * self.mean_,
scale=self.sigma_)
cdf = truncnorm.cdf(x_axis,
-b,
b,
loc=a * self.mean_,
scale=self.sigma_)
probability = probability / ((cdf[-1] - cdf[0]))
cdf = (cdf - cdf[0]) / ((cdf[-1] - cdf[0]))
return probability, cdf, x_axis
def get_probabilities(self):
if self.probabilities is None:
avg, std = self._get_mean_std_percentage()
if self.model_std is not None:
std = (self.model_std + std) / 2
if self.flat_uncertainty is not None:
std += self.flat_uncertainty
self.properties['Average interest'] = avg
self.properties['Std interest'] = std
if self.limit_std is not None:
std = self.limit_std_dev(std)
self.properties['Limited std'] = std
if self.prop_domain is not None:
a, b = (self.prop_domain[0] -
avg) / std, (self.prop_domain[1] - avg) / std
self.probabilities = truncnorm.cdf([0, 1],
a=a,
b=b,
loc=avg,
scale=std)
else:
self.probabilities = norm(avg, std).cdf([0, 1])
return self.probabilities
def log_prob(self, value):
if self._validate_args:
self._validate_sample(value)
return super(TruncatedNormal, self).log_prob(
self._to_std_rv(value)) - self._log_scale
if __name__ == '__main__':
from scipy.stats import truncnorm
loc, scale, a, b = 1., 2., 1., 2.
tn_pt = TruncatedNormal(loc, scale, a, b)
mean_pt, var_pt = tn_pt.mean.item(), tn_pt.variance.item()
alpha, beta = (a - loc) / scale, (b - loc) / scale
mean_sp, var_sp = truncnorm.stats(alpha,
beta,
loc=loc,
scale=scale,
moments='mv')
print('mean', mean_pt, mean_sp)
print('var', var_pt, var_sp)
print('cdf',
tn_pt.cdf(1.4).item(),
truncnorm.cdf(1.4, alpha, beta, loc=loc, scale=scale))
print('icdf',
tn_pt.icdf(0.333).item(),
truncnorm.ppf(0.333, alpha, beta, loc=loc, scale=scale))
print('logpdf',
tn_pt.log_prob(1.5).item(),
truncnorm.logpdf(1.5, alpha, beta, loc=loc, scale=scale))
print('entropy', tn_pt.entropy.item(),
truncnorm.entropy(alpha, beta, loc=loc, scale=scale))
def _cdf(self, x, a, b, mu, sigma):
return truncnorm.cdf(x, a, b, loc=mu, scale=sigma)
""" from scipy.stats import truncnorm import matplotlib.pyplot as plt fig, ax = plt.subplots(1, 1) a, b = 0, np.inf mean, var, skew, kurt = truncnorm.stats(a, b, moments='mvsk') x = np.linspace(truncnorm.ppf(0, a, b), truncnorm.ppf(0.99, a, b), 100) ax.plot(x, truncnorm.pdf(x, a, b), 'r-', lw=5, alpha=1, label='truncnorm pdf') mean = 2 rv = truncnorm(a, b) ax.plot(x, rv.pdf(x), 'k-', lw=2, label='frozen pdf') vals = truncnorm.ppf([0.1, 0.1, b], a, b) np.allclose([0.0001, 1.5, 2], truncnorm.cdf(vals, a, b)) r = truncnorm.rvs(a, b, size=1000) ax.hist(r, density=True, histtype='stepfilled', alpha=1) ax.legend(loc='best', frameon=False) plt.show() plt.plot(r) mu, sigma = 10, 0.1 s = np.random.normal(mu, 0, 1000)
def get_truncated_lognormal_example_exact_quantities(lb, ub, mu, sigma):
f = lambda x: np.exp(x).T
#lb,ub passed to truncnorm_rv are defined for standard normal.
#Adjust for mu and sigma using
alpha, beta = (lb - mu) / sigma, (ub - mu) / sigma
denom = normal_rv.cdf(beta) - normal_rv.cdf(alpha)
#truncated_normal_cdf = lambda x: (
# normal_rv.cdf((x-mu)/sigma)-normal_rv.cdf(alpha))/denom
truncated_normal_cdf = lambda x: truncnorm_rv.cdf(
x, alpha, beta, loc=mu, scale=sigma)
truncated_normal_pdf = lambda x: truncnorm_rv.pdf(
x, alpha, beta, loc=mu, scale=sigma)
truncated_normal_ppf = lambda p: truncnorm_rv.ppf(
p, alpha, beta, loc=mu, scale=sigma)
# CDF of output variable (log truncated normal PDF)
def f_cdf(y):
vals = np.zeros_like(y)
II = np.where((y > np.exp(lb)) & (y < np.exp(ub)))[0]
vals[II] = truncated_normal_cdf(np.log(y[II]))
JJ = np.where((y >= np.exp(ub)))[0]
vals[JJ] = 1.
return vals
# PDF of output variable (log truncated normal PDF)
def f_pdf(y):
vals = np.zeros_like(y)
II = np.where((y > np.exp(lb)) & (y < np.exp(ub)))[0]
vals[II] = truncated_normal_pdf(np.log(y[II])) / y[II]
return vals
# Analytic VaR of model output
VaR = lambda p: np.exp(truncated_normal_ppf(p))
const = np.exp(mu + sigma**2 / 2)
# Analytic VaR of model output
CVaR = lambda p: -0.5 / denom * const / (1 - p) * (erf(
(mu + sigma**2 - ub) / (np.sqrt(2) * sigma)) - erf(
(mu + sigma**2 - np.log(VaR(p))) / (np.sqrt(2) * sigma)))
def cond_exp_le_eta(y):
vals = np.zeros_like(y)
II = np.where((y > np.exp(lb)) & (y < np.exp(ub)))[0]
vals[II] = -0.5 / denom * const * (erf(
(mu + sigma**2 - np.log(y[II])) / (np.sqrt(2) * sigma)) - erf(
(mu + sigma**2 - lb) / (np.sqrt(2) * sigma))) / f_cdf(y[II])
JJ = np.where((y >= np.exp(ub)))[0]
vals[JJ] = mean
return vals
ssd = lambda y: f_cdf(y) * (y - cond_exp_le_eta(y))
mean = CVaR(np.zeros(1))
def cond_exp_y_ge_eta(y):
vals = np.ones_like(y) * mean
II = np.where((y > np.exp(lb)) & (y < np.exp(ub)))[0]
vals[II] = -0.5 / denom * const * (erf(
(mu + sigma**2 - ub) / (np.sqrt(2) * sigma)) - erf(
(mu + sigma**2 - np.log(y[II])) /
(np.sqrt(2) * sigma))) / (1 - f_cdf(y[II]))
JJ = np.where((y > np.exp(ub)))[0]
vals[JJ] = 0
return vals
ssd_disutil = lambda eta: (1 - f_cdf(-eta)) * (eta + cond_exp_y_ge_eta(-eta
))
return f, f_cdf, f_pdf, VaR, CVaR, ssd, ssd_disutil
# Plots
# plt.figure(1)
# plt.plot(x,truncnorm.pdf(x, a, b, loc=mu, scale=sigma),'b',label='normpdf')
# plt.legend()
# plt.figure(2)
# plt.plot(x,Weightedpdf(x,mu,sigma,a,b),'r',label='pdf/x**2')
# plt.legend()
# plt.figure(3)
# plt.plot(x,awPDFDistribution(x,mu,sigma,interval,a,b),'r',label='awPDF')
# plt.legend()
plt.figure(4)
plt.plot(x,
truncnorm.cdf(x, a, b, loc=mu, scale=sigma),
'b-',
label='Normal Distribution')
plt.plot(x,
Phi(x, mu, sigma, x, interval, a, b),
'b--',
label='Area Weighted Normal Distribution')
# ''' subplots
f, (ax1, ax2) = plt.subplots(1, 2)
ax1.plot(x,
truncnorm.cdf(x, a, b, loc=mu, scale=sigma),
'b-',
label='ND of (20,8)')
ax1.plot(x,
Phi(x, mu, sigma, x, interval, a, b),
for i in range(ndim):
arr_1[:, i] = sampler._rvs[str(i)].flatten()
arr_2[:, i] = samplerEnsemble._rvs[str(i)].flatten()
colors = ["black", "red", "blue", "green", "orange"]
plt.figure(figsize=(10, 8))
for i in range(ndim):
s = np.sqrt(cov[i][i])
### get sorted samples (for the current dimension)
x_1 = arr_1[:, i][np.argsort(arr_1[:, i])]
x_2 = arr_2[:, i][np.argsort(arr_2[:, i])]
### plot true cdf
plt.plot(x_1,
truncnorm.cdf(x_1, llim, rlim, mu[i], s),
label="True CDF",
color=colors[i],
linewidth=0.5)
# NOTE: old mcsampler stores L, mcsamplerEnsemble stores lnL
L = sampler._rvs["integrand"]
p = sampler._rvs["joint_prior"]
ps = sampler._rvs["joint_s_prior"]
### compute weights of samples
weights_1 = (L * p / ps)[np.argsort(arr_1[:, i])]
L = samplerEnsemble._rvs["integrand"]
p = samplerEnsemble._rvs["joint_prior"]
ps = samplerEnsemble._rvs["joint_s_prior"]
### compute weights of samples
weights_2 = (L * p / ps)[np.argsort(arr_2[:, i])]
y_1 = np.cumsum(weights_1)
def MLE_X_new_omega(self,low_support,high_support,threshold,x2nd):
'''
this is old version that use the probability of
Prob(Xi in [xi_low, xi_up] | Omega_it xj in [xj_low,xj_up])
'''
self.setup_para(0)
mu=self.MU[-2]
sigma=self.SIGMA2[-2,-2]**0.5
low_support = low_support.reshape(self.N,1)
high_support = high_support.reshape(self.N,1)
density_2nd = truncnorm.pdf((x2nd-mu)/sigma,(threshold[-2]-mu)/sigma,10)
prob_1st = 1 - truncnorm.cdf((low_support[-1]-mu)/sigma,(threshold[-1]-mu)/sigma,10)
with np.errstate(divide='raise'):
try:
log_Prob = np.log(density_2nd) + np.log(prob_1st)
except Exception as e:
print('-----------------------------------------------')
print('0 in log at {} bidders with {} reserve price'.format(self.N,threshold[0]))
print(low_support.flatten())
print(high_support.flatten())
print("density_2d: {} | prob_1st: {}".format(density_2nd[0],prob_1st[0]))
log_Prob = np.nan
return log_Prob
if self.N>2:
for i in range(self.N-2):
temp_low =low_support[i+1:]
temp_high =high_support[i+1:]
temp_low =np.append(temp_low,threshold[i])
temp_high =np.append(temp_high,10)
[x_v,w_v]=self.GHK_simulator(i,temp_low,temp_high,2)
# last column is what we need
#
x_flag1 = x_v[0] >= low_support[i]
x_flag2 = x_v[0] <= high_support[i]
check_flag_v1 = x_flag1*x_flag2*1
# calculate the prob
nominator = np.sum(check_flag_v1*w_v)
denominator = np.sum(w_v)
with np.errstate(divide='raise'):
try:
log_Prob = log_Prob + np.log(nominator)-np.log(denominator)
except Exception as e:
print('-----------------------------------------------')
print(e)
print('0 in log at {} bidders with {} reserve price for bidder {}'.format(self.N,threshold[0],i))
print(low_support.flatten())
print(high_support.flatten())
print("density_2d: {} \t| prob_1st: {} \t| nominator: {} \t| denominator: {} \t| ".format(density_2nd[0],prob_1st[0],nominator,denominator))
log_Prob = np.nan
return log_Prob
def MLE_X_new_karl(self,low_support,high_support,threshold):
'''
In Karl's suggestion, conditional each Omgea_i, I can get conditional distribution for
each bidder i. Then I can calculate the probability that xi is under the lower and upper
bound
Prob_Xi (Xi in [X_low, X_up] | Xi > gamma)
Prob_xi( xi_low < xi < xi_up | Omega_it)
ignore the second highest bid first
For each xi I don't even need the truncated GHK simulator
Here I am doing miniziation to estimate the prob that outside the support
'''
self.setup_para(0)
mu=self.MU
sigma=np.diag(self.SIGMA2)**0.5
old_low=np.copy(low_support)
old_high=np.copy(high_support)
flag=low_support[:-2]>high_support[:-2]
high_support[:-2]=(1-flag)*high_support[:-2]+flag*high_support[-2]
flag=high_support[:-2]>high_support[-2]
high_support[:-2]=(1-flag)*high_support[:-2]+flag*high_support[-2]
low_support = low_support.flatten()
high_support = high_support.flatten()
# Notice that P_Xi (Xi in [X_low, X_up] | Xi > gamma)
# from i=1, 3,4,....
# minimize ignore the second highest
norm_threshold = (threshold[-1]-mu[-1])/sigma[-1]
norm_support = (low_support[-1]-mu[-1])/sigma[-1]
prob_1st=truncnorm.cdf(norm_support,norm_threshold,15)
with np.errstate(divide='raise'):
try:
log_Prob = np.log(1+prob_1st)
except Exception as e:
print('-----------------------------------------------')
print('0 in log at {} bidders with {} reserve price'.format(self.N,threshold[0]))
print(low_support.flatten())
print(high_support.flatten())
print("density_2d: {} ".format(prob_1st[0]))
log_Prob = np.nan
return log_Prob
if self.N > 2:
for i in range(0,self.N-2):
norm_threshold = (threshold[i]-mu[i])/sigma[i]
norm_low_supp = (low_support[i]-mu[i])/sigma[i]
norm_high_supp = (high_support[i]-mu[i])/sigma[i]
Prob_temp1 = truncnorm.cdf(norm_low_supp,norm_threshold,15,mu[i],sigma[i])
Prob_temp2 = 1 - truncnorm.cdf(norm_high_supp,norm_threshold,15,mu[i],sigma[i])
with np.errstate(divide='raise'):
try:
log_Prob = log_Prob + np.log(1+Prob_temp1) + np.log(1+Prob_temp2)
except Exception as e:
print('-----------------------------------------------')
print(e)
print('0 in log at {} bidders with {} reserve price for bidder {}'.format(self.N,threshold[0],i))
print(low_support.flatten())
print(high_support.flatten())
print("prob_1st: {} \t| low: {} \t | upp: {} \t ".format(prob_1st,Prob_temp1,Prob_temp2))
log_Prob = np.nan
return log_Prob
def normal_truncated_cdf(self, xvalue):
return truncnorm.cdf(
xvalue, (self.lower_limit - self.prior_estimate) / self.spread,
(self.upper_limit - self.prior_estimate) / self.spread,
loc=self.prior_estimate,
scale=self.spread)